Changelog archive — through 2026-08-13

Entries older than the current wave, moved out of the root CHANGELOG.md on 2026-08-25 so the live file stays readable. Nothing was edited: this is the same text, newest-first, continuing where CHANGELOG.md stops.

Looking for a decision’s rationale? The changelog records what changed and why on the day. Rules that outlived their entry live in ../AGENTS.md and windows-build-invariants.md; symptoms live in failure-modes.md. If you find a rule that exists only here, promote it.

2026-08-13 - VALIDATING REBUILD COMPLETE: :latest-cross re-shipped, 3 arches, every keeper change proven by a real build

• :latest-cross rebuilt base→manifest and pushed — manifest digest sha256:8d1538b1…, 3-arch (amd64 35c1f1df, arm64 e677e4f8, riscv64 5002954385), all runtime smokes 0 failures, [INFO] Cross chain complete. This validates the entire 2026-08-12 “fix everything” pile against a real build — the thing static gates (lint/tests/copy-coverage) structurally cannot prove.

• Keeper changes proven live: cpython-dev-packages table (base install-os-packages + toolchain build_python both built clean), BS1/BS2/BS3/ BS3b (sdk self-contained NEEDED walk clean, dead TVM RUN gone), S4 mounts, setup_gi 8-phase split, R1-R5, D4/P4 (BUNDLED: libtensorflow.so.2 + transitive libtensorflow_framework.so.2), wheel_family, csound patch-first, C1/C2 + NDK cache, and the recovery fixes (Fix #10 sympy backfill fired on riscv64; STV1 genai policy: not installed (documented riscv64 skip; policy, not drift) → VERSION-ASSERT PASS).

• Rebuild-surfaced fixes (latent over-strictness / infra the rebuild flushed out, none in the keeper changes): reverted TG1/TG3/TG7 (build-speed tweaks that failed a real build — libLLVMSupportLSP install, missing clang-tblgen, lazy-mount closure — all re-backlogged with redo recipes); BS3b libclang glob (libclang.so.2*→libclang*.so.2*, its own assert caught it); ORT-WEB1 (onnx-web hard-fail on a transient npm/GitHub error → optional/ship-without); BS5-followup (riscv64 Node major-assert die→warn, unavoidable ports lag).

• Infra: disk exhaustion twice (freed 367G via slug prune + nerdctl builder prune; android-riscv64 ld.lld Bus error was OOD, not code) and a corrupt local cache slug → switched resumes to CROSS_NO_LOCAL_CACHE_EXPORT=1.

2026-08-12 (afternoon) - “fix everything” wave 2 + a mass-agent-failure recovery

• Landed & validated (12 more Batch-2/legacy items, all gates green): STV1 root fix (smoke-torch-venv carries the genai-on-riscv64 policy itself; STV_REQUIRE_GENAI=1 re-arms; host exemption now TRANSITIONAL); the wheel-family classifier (one wheel_family() fn, 4 sites, 21-name equivalence battery); csound patch-first; MAKEINFO=true (GCC log noise); opencv cross-dep presence WARN; TG3+TG7 (ONE LLVM compile/arch + minimal host-LLVM profile after a zero-consumer audit); D4+P4 (elf_needed_sonames/elf_unresolved_needed primitives + NEEDED-driven SDK-lib bundling); setup_gi_cross_wrappers split into 8 phases; S4 (per-file install-deps mounts — COPY-fallback-safe); NDK shared android-sdk download cache; BS1/BS2/BS3/BS3b (sdk COPY closure, dead TVM RUN removed, real ELF verification, self-contained amd64 llvm prefix); TS3 + partial TS2 (see below). 22 suites / 317 assertions, copy-coverage 9/9, verify-critical-fixes 0 failures.

• TS3 done / TS2 partial: new 01-core/cpython-dev-packages.sh is the single dev-package<->extension table; both the host list (package-lists.sh) and the cross list (build_python.sh) derive from it, ending the desync that caused the 2026-08-09 libsqlite3-dev incident. The required/optional fatality split + lib-dynload asserts remain (need a cross rebuild).

• Recovery: a Fable-5 session limit killed 8 subagents mid-edit. Post- mortem + consolidation on Opus 4.8: (1) found and fixed a real double build-break — the dying TS agent left build_python.sh (and package-lists.sh) sourcing the new table with no mount and no COPY fallback; added the mounts in Dockerfile.base + Dockerfile.toolchain and wired build_python to the table. (2) Reverted TG1 (lazy-dispatch + 17-line mount trim, killed mid-trim) to its shipped HEAD state — a toolchain RUN has no COPY fallback so a missed mount = a multi-hour break, and copy-coverage.py can’t see source_module-by-name gaps; re-opened with a “needs per-RUN audit + real rebuild” note. (3) Confirmed the run_priv, D3/P5-smoke-scaffold, and NVIDIA-sweep agents landed nothing (no partial state). Built a differential source_module-vs-mount auditor to prove no other new closure gaps.

2026-08-12 (day) - backlog execution wave: Batch 0 CLOSED + 17 Batch-2 items landed in one opened closure window

• Batch 0 closed (post-ship window): canonical buildkitd config APPLIED + daemon restarted — gckeepstorage=500GB restored (the regression), NEW max-parallelism=4 (bounds the intra-arch DAG overcommit behind the 2026-08-10 opencv OOM), BUILDKIT_STEP_LOG_MAX_SIZE=50MiB live; verify-host-config reports drift-free. S1 salvage-cache-export: on non-transient build failure cross-stage-build.sh now sweeps the Dockerfile’s named stages with –target rebuilds (completed subtrees cache-hit in seconds and their local cache export finally lands; SALVAGE_CACHE_EXPORT=0 opts out, per-target timeout, stops after 2 consecutive failures). O3: orchestrator emits ${LOG_DIR}/chain-status.json (atomic, best-effort) at stage start/ok/fail with captured digest pins. B5: smoke-runtime-image main() split into 23 per-check functions — validated by before/after runs against the live amd64 wrapper (verdict lines byte-identical). CROSS_CACHE_MAX_GB=250 total-size cap added as disk-guard phase 2. CCACHE_MAXSIZE investigate RESOLVED with data (live limit already 30GB, 17.6% hits/129k — key churn, not size; → TG2).

• Batch-2 window opened deliberately (S1’s 01-core edit busts the toolchain/sdk/package whole-dir COPY caches → next chain run pays ONE full rebuild) and 17 items pulled into it: R1-R5 (ffmpeg-TF extraction/bundle guards, downloads no-HEAD split, torch/vision toolchain-file guards, opencv install stderr), C1/C2 (Cerbero apt fatal; android-sdk rc-only success + fatal licenses + deduped package list + NDK postcondition), BS4-BS6 (7 stale fallback literals → :? half-load guards, loud riscv64 node unpin + major assert, requested-but-absent package summary + must-have postcondition), TG2/TG4 (ccache stats blocks in the compiler stage; llvm-cross cmake –install –strip), XC4-XC7 (ONNX_PACKAGE/ PYTORCH_EXTRA actually wired ARG→ENV→forward + KNOWN_DEAD pruned to empty, smoke-vulkan comment truth, Dockerfile.package TARGETARCH- parameterized BASE_IMAGE default, VULKAN_VERSION hand-forward deleted + new no-hand-forward checker rule), D2 (retag_directory_wheels promoted to 01-core, 3 drifted copies replaced, copy-coverage green), A3 (abseil-headers out of the eager loop).

• SCR1 retired honestly: the two remaining strings-pipelines are deliberate (embedded DEB metadata beats –version, which apt’s libclang-cpp shadowing falsifies; executed fallback covers dylib-blind binaries) — not the fix-#8 class. Rider landed: test-invocation-lints now asserts force-reinstall ⇒ –no-deps.

• Pre-existing IFS comma-split in python_uv.sh fixed (test-ifs-safety was red). Gates: preflight rc=0, 23 suites green (312+ assertions), verify-script-copy-coverage green, verify-critical-fixes rc=0.

2026-08-12 - :latest-cross SHIPPED: 3-arch manifest pushed after fixes #6-#10; the runtime-lane gate saga closes at 10 fixes / 8 runs

• GOAL ACHIEVED: ghcr.io/kataglyphis/kataglyphis_beschleuniger:latest-cross manifest pushed (digest sha256:8466ba39d395…994a), verified 3-arch (amd64 2be255c2…, arm64 37ca9c0c…, riscv64 a1a3038a…); all per-arch runtime smokes 0 failures; [INFO] Cross chain complete. (run #8, log chain-runtime-fix10-20260812-040248.log).

• Fix #10 — riscv64 torchvision No module named 'sympy' (run #7’s only failure; fix-#7 collateral): on riscv64 torch ships as LOCAL wheels, so the torch backend extra is never requested from uv sync and the lock graph omits torch’s own runtime deps; the old full-deps --force-reinstall dragged them in as the exact side effect --no-deps rightly killed. Fix: post-sync import-probe backfill in reconcile_local_wheels (assemble-torch-app.sh) over sympy/mpmath/networkx/jinja2/markupsafe/ filelock/fsspec/typing-extensions, each installed --no-deps; live run backfilled exactly the 5 missing leaves, amd64/arm64 probes all satisfied (no-op). riscv64 app-wheel smoke then passed on-target incl. torchvision ops.nms.

• riscv64 genai exemption validated in anger: the host-side smoke-runtime-image exemption converted the expected genai absence to PASS while every other mismatch stays fatal. Researched upstream: genai has NO riscv64 support (no PyPI wheel in any version, no riscv CI, issue #594 closed “not planned”) — our skip is upstream-consistent; optional self-build filed as backlog GEN1, root fix remains STV1.

• Diagnosis gotcha recorded: smoke-torch-venv prints its OWN === Results: N failure(s) === banner inside the captured-and-echoed assert output; an outer-looking “1 failure(s)” mid-suite is the echoed INNER summary, not the host verdict (which may arrive minutes later under QEMU). Filed in the session memory to prevent re-diagnosis.

• Backlog: +SCR1 (two more strings-scrape version checks: setup-package-image.sh:150, validate-compilers.sh:48 + a force-reinstall/ no-deps pairing lint), +GEN1. Batch 0 (post-chain window) is now UNLOCKED.

2026-08-11 (night) - runs 19+20 decode the shim-4 flip-flop: patch_file_content silently NO-OPED; shim 4 rebuilt as REMOVE_RECURSE; run 21 launched

• Run 19 (forensic line): the container READS the current patch file (“5801 chars; examples-drop block present: True”) - run 18’s miss was a stale-context one-off. Run 19 then died at a TRANSIENT dep download (“cmake -E tar: ZIP decompression failed (-5)” at 92%) BEFORE reaching gemma3 - which masked the real story.

• Run 20 brought the Protobuf death BACK and completed the picture: the shim-4 STATUS message had printed all along (hidden by buildkit’s merged log lines), but patch_file_content’s string-replace NEVER MATCHED (the helper prints unconditionally - a no-op looks identical to success), so gemma3 kept configuring. Lesson recorded: never trust an unconditional-message patch helper; verify by absence of the symptom.

• Shim 4 rebuilt drift-immune: file(REMOVE_RECURSE "${LITERT_SRC_DIR}/tensor/examples") + self-verifying EXISTS check (loud WARNING if the tree survives) - upstream’s own if(EXISTS .../CMakeLists.txt) guard then skips both example dirs. Run 21 launched. media-core meanwhile at SIX consecutive fully-green runs (ONNX 10x bare links).

2026-08-11 - knowledge placement: a rule, an index, and a consumer AGENTS.md skeleton

Code reuse has a rule (the two-consumer test). Documentation did not, and it showed: the Dev Drive filter command existed in three places - windows-builds.md here plus Inference-Engine’s AGENTS.md and docs/source/platforms.md - and all three were wrong the same way while windows-container-build-performance.md had it right AND warned about that exact mistake. Restating produced three broken copies.

• AGENTS.md § Where does knowledge belong? states the rule: would this still be true in a different project? Yes -> here, and the consumer links. No -> the consumer writes it out. Most topics split down the middle.

• docs/INDEX.md (new) maps topic -> owning document. Consumers link one hop through it, so reorganising docs here means editing that page rather than hunting links across seven repositories.

• shared/templates/AGENTS.md.template (new) is the consumer skeleton, with section 2 (“what ANTfrastructure owns”) deliberately links-only.

Explicitly NOT added: a lint that fails consumer docs for mentioning wcifs or --isolation process. A keyword cannot tell restating from applying - Inference-Engine’s ASan notes discuss image-level runtimes legitimately, because which runtime a Flutter/COM app survives is a property of that app.

2026-08-11 - the Python ci_*.sh drivers were unrunnable; both consumers had reimplemented them

02-toolchain/python/ci-common.sh sourced ../01-core/python_uv.sh. That file sits one level deeper than 02-toolchain/bootstrap.sh, where ../01-core IS correct, so it resolved to 02-toolchain/01-core/python_uv.sh - a path that has never existed. Every driver in that directory therefore died on its first helper call with detect_workspace: command not found.

Which explains the duplication it was hiding: Kataglyphis-Orchestr-ANT-ion and Kataglyphis-WebDavClient each carried a full local reimplementation of all four drivers - 339 and 383 lines - because the shared ones could not run. Both now delegate; 722 lines became 120.

Also picked up from the consumer copies while collapsing them:

• ci_tests.sh now guards the dependency SYNC for experimental interpreters the way it already guarded venv creation. A free-threaded build with one unbuildable wheel used to fail the whole matrix.

Not upstreamed, per the two-consumer rule: WebDavClient’s patchelf install (one consumer) stays in its wrapper.

2026-08-11 (midday) - run 17: shim 3 proven (litert @bazel-truth fetched, API-skew gone), configure died on litert’s UNCONDITIONAL example projects -> shim 4; run 18 launched

• Run 17 (survived a harness restart mid-flight; driver kept orchestrating): media-core fully green for the THIRD consecutive run (ONNX 7th bare link). litert-lm: pin bump applied and fetched (“Already at requested ref: 3cb830ad”), the SetEnableYNNPack/RunAsync API-skew errors are GONE - the failure moved into the NEW litert commit itself: @3cb830ad adds tensor/examples/{segmentation,gemma3} guarded only by file-existence (no option), and gemma3 does find_package(Protobuf REQUIRED) -> litert_external CONFIGURE death. Shim 4 (litert-patcher-winfix.cmake): drop both add_subdirectory(examples/…) calls - examples are not in litert-lm’s target map. google-ai-edge cmake-staleness set now at FOUR findings.

2026-08-11 - coverage.sh: optional HTML report for the llvm backend

coverage_llvm_report gained COVERAGE_LLVM_HTML_DIR, which additionally emits the browsable llvm-cov show report.

Optional because the two consumers genuinely differ: BeschleunigerBallett only feeds Codecov from the JSON export, while Kataglyphis-Cpp-Inference publishes the HTML next to its docs. Without it that second consumer could not use this function at all - it had to keep a private llvm-profdata/llvm-cov pipeline, which is precisely the duplication this library exists to remove. It now drives both backends (gcovr for GCC, llvm-cov for Clang) from here.

2026-08-11 - shared config: canonical was the stale copy, and -Ignore never worked via -File

Resolving Kataglyphis-Cpp-Inference’s four-file drift report turned the conclusion around: it was not deviating, it was AHEAD, and three canonical files were wrong for every C++ consumer.

• .clang-format: Standard: c++20 -> c++23. BeschleunigerBallett sets CMAKE_CXX_STANDARD 23; canonical had never been updated.

• .pre-commit-config.yaml: the clang-format files: regex omitted .ixx, so BeschleunigerBallett’s 63 module interface units were never formatted.

• .clang-tidy: misc-include-cleaner disabled - noise on module-using code.

All three written out to BeschleunigerBallett, which is back in sync. Cpp-Inference keeps three genuine overrides (documented in shared/config/README.md with the reason for each): its .clang-tidy answers the “clang-tidy sees import without BMIs” problem by suppressing clang-diagnostic-error where BeschleunigerBallett answers it by skipping module TUs entirely, its gcovr.cfg excludes follow its layout, and its pre-commit runs extra hooks. Its .clang-format is no longer an override.

-Ignore was broken via pwsh -File - the invocation the README itself shows. Under -File every argument is a plain string, so -Ignore a,b,c bound as ONE element "a,b,c", matched no file name, and the documented escape hatch silently did nothing. It now splits comma-separated values, so -File and -Command behave identically, and an entry that names no managed file throws instead of being quietly dropped.

2026-08-11 - WindowsOnnx.Common restored (silently-lost consumer dependency)

Widening the sweep to every Kataglyphis repo turned up a seventh consumer - Kataglyphis-Cpp-Inference, reached through Inference-Engine -> NativeInferencePlugin -> native/KataglyphisCppInference - and its Build-Windows.ps1 imports two modules that no longer exist here:

• WindowsLogging.Common was FOLDED into WindowsBuild.Common (b391a1d), so the consumer just drops the import; the Write-BuildLog* wrappers it actually uses are exported from there.

• WindowsOnnx.Common was DELETED outright (f0d12ff) while the consumer still calls Get-OnnxPackageLayout. Restored here. This is the third instance of the same blind spot (Sync-BuildArtifacts, the WindowsToolchain/Flutter/CodeQL trio, now this): a “zero callers” sweep can only see this repo, and consumers pin a submodule commit, so they neither break at delete time nor appear in the sweep.

  This one was the worst of the three because the call site sits inside a try/catch: losing it did not fail the build, it silently skipped the ONNX include/lib wiring.

  One substitution against the deleted original: it resolved paths through Resolve-ContainerNormalizedPath, itself a pure pass-through to Resolve-NormalizedPath and since removed as well. The restored module calls Resolve-NormalizedPath (WindowsScripts.Shared) directly.

Tests: 480/480 green.

2026-08-11 - Codacy removed org-wide

Codacy is dropped from every Kataglyphis repository at the owner’s request. Here that means .codacy/ (adopted only hours earlier in the pass-2 sweep below - it turned out the direction of travel was removal, not consolidation), the gitignored .github/instructions/codacy.instructions.md, and the .codacy entries in .gitignore / .dockerignore.

Nine repositories were swept: ANTfrastructure, BeschleunigerBallett, Inference-Engine, Orchestr-ANT-ion, RustProjectTemplate, WebDavClient, jotrockenmitlocken, Kataglyphis and Cpp-Inference (plus DocumANTation’s .dockerignore). No README badges and no CI steps referenced Codacy, so nothing else had to change. The pass-2 entry below is left as written: it is history, and it records why .codacy/ briefly lived here.

2026-08-11 - consumer-duplication sweep, pass 2: shared config completed, .codacy adopted

Second pass over the four consumers. The PowerShell, bash and CI surfaces came back clean; what was left was shared configuration, and two live defects.

• .codacy/ (new: cli.sh, codacy.yaml) — Kataglyphis-Orchestr-ANT-ion and Kataglyphis-RustProjectTemplate both symlink .codacy/cli.sh to ../ExternalLib/Kataglyphis-ANTfrastructure/.codacy/cli.sh, and this repo had no .codacy/ at all: both symlinks were dangling. Adopted from Inference-Engine’s real copy (Codacy’s stock CLI bootstrap, nothing project-specific in it), which makes the two symlinks resolve and lets the third consumer stop carrying 173 lines of its own.

• Sync-SharedConfig.ps1 no longer misdirects on a missing canonical file. If a name in $names has no file next to the script, -Write dies inside Copy-Item with a bare “path not found” and -Check reports the file as missing in the consumer - both sending you to the wrong repository. It now throws a message naming this directory. Hardening only: all four canonical files are present and -Check reports BeschleunigerBallett fully in sync.

• README documents the standing override: the canonical set is C++-oriented, so a Python consumer owns its own .pre-commit-config.yaml (ruff, not clang-format) and passes -Ignore .pre-commit-config.yaml. That is Orchestr-ANT-ion, and its difference is a deliberate override, not drift.

Not moved, deliberately: Inference-Engine’s packaging-common.sh (deb / AppImage / flatpak bundling, 502 lines). No second consumer packages desktop bundles, so it fails the two-consumer test and stays where it is.

2026-08-11 (mid-morning) - run 16: litert-lm shims 1+2 PROVEN (past run-15’s deaths), failure moved to UPSTREAM STALENESS #3 (litert pin) -> shim 3, run 17 launched

• Run 16: second consecutive fully-green media-core (ONNX 6th bare link, OpenCV 3rd, FFmpeg 2nd, GenAI 2nd). media-litert phase 1 (LiteRT v2.1.6) green in ~11 min; LiteRT-LM applied BOTH staleness shims cleanly (proto list + absl 20260526) and got PAST run-15’s compile deaths - then died at 1879 s on the NEXT layer: their cmake litert pin (fb16353a, ‘#Updated on 2026-03-24’) predates the LiteRT APIs 0.15.0’s own executor calls (CpuOptions::SetEnableYNNPack missing, CompiledModel::Run/RunAsync signature mismatches; plus a kv_cache.cc reference into a path that no longer exists - suspected same-pin fallout). The bazel WORKSPACE pins LITERT_REF=3cb830ad for this tag; shim 3 bumps their cmake pin to the bazel truth. Staleness count for google-ai-edge/LiteRT-LM v0.15.0’s cmake lane: THREE (protos, absl, litert) - issue-worthy as a set.

• Tree note: two new files + a PSSA warning appeared in windows (WindowsTesting.Common.psm1 empty-catch) from outside this session - left untouched.

2026-08-11 - consumer-duplication sweep: two modules upstreamed, uv lock-retry, Resolve-BuildModule template

Audited all four consumers (BeschleunigerBallett, Inference-Engine, Orchestr-ANT-ion, RustProjectTemplate) for code that duplicates what this repo already ships, and for local code that more than one of them needed.

• windows/scripts/modules/WindowsTesting.Common.psm1 (new) — upstreamed from BeschleunigerBallett’s vendored Scripts/Windows/modules copy. Test-exe discovery (Resolve-TestExecutable), ctest driving (Invoke-CtestDiscoveredTests), scoped ASAN_OPTIONS (Invoke-WithAsanOptions) and ASan-runtime discovery (Get-AsanRuntimeDirs, Get-AsanRuntimeDll). Nothing in it was project-specific. Second consumer proving it belongs here: Inference-Engine’s Start-Windows.ps1 had its own narrower copy of the runtime search that only ever matched the BuildTools SKU. The Visual Studio half now resolves through Get-MsvcToolsRoots (WindowsScripts.Shared) instead of globbing Program Files*\Microsoft Visual Studio\*\*, so Community/Professional/Enterprise are found too, the cold-boot vswhere race is retried, and the NEWEST toolset comes first. Get-AsanRuntimeDirs lost its mandatory-but-unused -BuildRoot; Resolve-TestExecutable gained -AdditionalRelativeDirectory in place of the hard-coded Test\{commit,compile,perf} probes.

• windows/scripts/modules/WindowsClang.Common.psm1 (new) — likewise upstreamed. clang-tidy driving, the sibling of WindowsFormatting.Common’s clang-format/cmake-format (separate because tidy needs a compile-commands database). Its two project-specific values are now parameters with the old values as defaults: -SourceSubdirectory (was a hard-coded Src) and -ModuleImportPattern (was a hard-coded import kataglyphis).

• Sync-UvProjectDependencies gained -RetryWithoutLocked and a -LogWarning sink. Orchestr-ANT-ion had re-implemented the entire function locally just to get the retry-without---locked fallback. Deliberately OPT-IN, not the default: --locked exists so CI fails on an un-regenerated lockfile, and defaulting the fallback on would turn a reproducibility gate into a no-op.

• shared/windows/templates/Resolve-BuildModule.ps1 (new) — the bootstrap that cannot be imported from here (it is what finds this submodule) now has a canonical template, plus a Resolve-BuildModulePath back-compat alias. All four consumers had written it independently and drifted: different search orders, different error text, one missing -Global.

• docs/adopting-in-a-new-project.md updated for the new modules and template.

2026-08-11 (morning) - run 15 fell in media-litert: LiteRT-LM v0.15.0’s cmake lane is DOUBLY stale upstream; two shims added, run 16 launched

• media-litert died at 1757 s with two independent missing-header classes, both root-caused to UPSTREAM cmake staleness at v0.15.0 (their cmake lane clearly lags their bazel truth and is not CI-covered): (1) the proto list in cmake/packages/litert_lm/CMakeLists.txt enumerates only the v0.13-era protos - embedding_metadata/embedding_model_type/executor_metadata/ litert_lm_metrics are MISSING while 0.15.0 sources include their pb.h (verified against the tag’s runtime/proto listing); (2) their absl pin 20260107.1 predates absl/status/status_macros.h, which 0.15.0’s own sources include (header exists from 20260526.0 - verified against abseil-cpp tags). Fixes in Build-LitertLmFromSource.ps1, port style: proto-list append after token.proto + absl GIT_TAG bump to the repo-wide ABSEIL_VERSION (scope: absl_external only; tflite/litert fetch their own). Fourth genuine upstream bug of the night (sccache, opencv, 2x litert-lm) - upstreamable to google-ai-edge/LiteRT-LM.

2026-08-11 (morning) - run 15: MEDIA-CORE FULLY GREEN for the first time since the 2026-08-03 bump

• The whole media-core branch is green end-to-end: ONNX (5th consecutive bare-CUDA link), OpenCV (2nd green, patches 001-004), FFmpeg n9.0 (.version pre-generation PROVEN: configure 73 s, .pc guard passed, vertex committed), and ONNX GenAI 0.15.2 (Neuland, green in 339 s; CUDA 80/86/89/90, nvcc bare via the opt-in gate). Remaining chain: media-litert + media-tvm branches, GStreamer merge, torch, final.

• New P0 discovered along the way (backlog item 34): stage re-exports mint fresh manifest digests every run (zero digest overlap between runs 14/15 even for byte-identical base/sdk/toolchain), so downstream FROMs cache-miss chain-wide — the partitioned chain currently has NO cross-run caching. Fix directions recorded (SOURCE_DATE_EPOCH clamp / skip-export on hit / FROM-by-digest). Plus item 35: a transient ~120-min stall in the ffmpeg stage’s VsDevCmd/scoop section (self-recovered, bracketed in the log).

2026-08-11 (night) - run 14: OpenCV GREEN FOR THE FIRST TIME (patch 004 proven); FFmpeg n9.0 died in the NEW .version machinery -> PowerShell pre-generation; run 15 launched

• OpenCV vertex COMMITTED green - first time ever since the 5.0.0 bump: the full patch cascade 001 (cmake) -> 002 (mlas C++) -> 003 (mlas skip, proven in run 13) -> 004 (dnn/ORT wchar, proven this run: compile sailed past run 13’s 435 s death point, 1862 TUs, ninja + link complete).

• Run-14 stats sealed the caching mystery (dossier 2.7): 1498 (ONNX) / 1862 (OpenCV) compile requests, 0 hits, and write errors == misses — EVERY cache write of every real vertex fails, both levels. The whole caching apparatus has been a silent no-op all along. Config-matrix probes exonerate every single-compile combination INCLUDING multilevel+cache- mount, leaving concurrency or mount history; the chain’s own error log (1498 plain-text lines on the -2 mount) is the final witness, readable once the mount unlocks.

• FFmpeg n9.0 (Neuland) died at our own .pc guard - correctly: since n9.0, .pc Version fields come from GENERATED libX.version files (library.mak -> ffbuild/libversion.sh awk/eval); under the Git-Bash port that chain emitted empty MAJOR/MINOR/MICRO -> ‘Version: ..’ in every .pc (the guard existed because the OLD version of this failure shipped silently once). Fix: Build-FfmpegFromSource.ps1 pre-generates all eight libX/libX.version files from the version(.major).h macros (LF, no BOM - make -includes them for LIBVERSION/LIBMAJOR = DLL naming), hard-throwing if the macros are not found. Run 15 resumes at FFmpeg.

2026-08-11 (early) - run 13: patch 003 PROVEN (MLAS passed), OpenCV died one layer deeper in an UPSTREAM OpenCV bug -> patch 004; run 14 launched

• Run 13’s OpenCV vertex applied 003 cleanly and got PAST the vendored MLAS entirely (run 12’s killer) - died at 435 s in modules/dnn/src/net_impl_backend.cpp:99: dnn’s ORT profiling call passes char* to Ort::SessionOptions::EnableProfiling, but ORTCHAR_T is wchar_t on Windows. Genuine upstream OpenCV 5.0.0 bug: the model-path call four lines below is properly #ifdef _WIN32-widened, the profiling call is not - upstream Windows CI never compiles dnn with ORT enabled. 004-dnn-ort-profiling-wchar.patch mirrors the existing widening (generated via the git-diff method, apply-check green; upstreamable to opencv/opencv as-is). Patch gate 12/12 OK, lint clean; run 14 resumes at OpenCV (ONNX again cached-green from run 13’s commit).

• Also this run: ONNX linked green bare a SECOND time (runs 5/12/13 vs wrapped 10/11 - the attribution A/B is now 5 runs deep).

2026-08-11 (early) - fault-attribution round: CUDA class sealed as sccache-side; L2 mystery attributed to OUR dufs task lifecycle; idle-timeout stats artifact fixed

• Attribution dossier (out/sccache-fault-attribution.md), built on the owner’s challenge “beweise, dass der Fehler bei mir oder bei sccache liegt”: Class 1 (nvcc crash + silent instantiation loss) = sccache-side beyond reasonable doubt (7/8 evidence rows closed; only the internal mechanism awaits the exact-TU replay). Class 2 (L2 never fed) = sccache EXONERATED by a three-phase probe FROM the chain image: raw PUT/GET OK, webdav-only write OK (+1 remote), multilevel READ+backfill OK (accidental cross-phase hit - comment nonces don’t survive preprocessing, round-2 nonce moved into a symbol), multilevel WRITE-THROUGH OK (+1 remote, +1 local, 0 errors). Leading OUR-side theory (2.8): dufs is an ONLOGON session-bound task - alive at launch (endpoint gate passes), killable by a mid-run logoff, and multilevel fails OPEN (policy l0) with zero L2 entries as the only symptom. Seal pending: error-log timestamps from the cache mount after run 13. Durable fix queued: dufs as a session-independent service + single-instance guard.

• All-zero end-of-vertex stats (runs 12+13) root-caused as a measurement artifact: the build server idle-exits (600 s default) during the long bare-CUDA tail, so the stats query talks to a fresh server. SCCACHE_IDLE_TIMEOUT=0 baked into both media Dockerfiles - run 14 delivers the first REAL per-vertex counters (and the async write-through tail no longer dies with the server).

• Run 13: ONNX vertex GREEN again (second consecutive bare-CUDA link); OpenCV solve running with the full new closure incl. patch 003 - MLAS verdict pending.

2026-08-10 (night, review round 2) - two-agent audit over the night’s diff: 1 confirmed design flaw + 3 plausible bugs fixed, 2 doc/config gaps closed

• CUDA launcher flipped to OPT-IN at the wiring site (confirmed find: build-onnx’s process-wide SCCACHE_NO_CUDA_LAUNCHER=1 leaked into later same-process stages on the classic lane while the BK lane kept wrapping OpenCV/GenAI CUDA through the miscompile-disqualified path - the lanes disagreed). Invoke-CmakeConfigure now adds the CUDA launcher only under SCCACHE_CUDA_LAUNCHER=1 (three-canary bar in the comment); the per-script opt-out env var is gone. Landed BEFORE run 13’s OpenCV solve, so its CUDA compiles run bare too.

• Test-Cache.ps1 per-run nonce: byte-identical source false-failed the writes>=1 assertion on the second-ever run against a warm WebDAV L2 (first compile = L2 hit, nothing stored). onnx-ninja.log rotation now Copy+Remove instead of Move-Item (cache mount is rename-hostile - probed for dirs, same wcifs family for files). A/B layer-lock: $disabled set when the disable is ISSUED, not after the 2 s post-state check - a slow driver teardown could previously strand the host dGPU-off with the finally-guard skipped.

• Doc/config gaps from the consistency agent: pending-cleanup tag list now names the orphaned pre-rename probe-build-copy[-heavy] tags (the findstr diag- one-liner cannot find them); new backlog item 33 (Test-PatchesApplyClean silently SKIPs patch dirs without a repo mapping); windows/buildkitd.toml comment updated for the -2 mount id incl. the transient 168h over-budget window of the orphaned -1 mount.

• Verified-clean list from the agents (for the record): $LASTEXITCODE through all new Tee pipelines, every new export/caller contract, StrictMode hardening, stall-guard job mechanics, patch-ladder idempotency, CI patch-drift trigger paths. Gates after fixes: lint 141 0/0 + AST traps clean, tests 457/457.

2026-08-10/11 - Linux lane: media 3/3 GREEN after 5 chain fixes; backlog program (6 sweep rounds, ~50 items executed); 2 new gates

• base->latest-cross chain rebuilt to media-3/3-pinned (amd64+arm64+riscv64 smokes all 0 failures; android/runtime/manifest in flight). Five fixes to get there, each found the hard way: (1) ffmpeg libtensorflow — the global --extra-libs=-ltensorflow broke FFmpeg’s EXECUTED configure sanity binary; fix = -lstdc++ only (the ONNX pattern) + bundle_tensorflow_runtime_lib copying the SDK .so out of the cache mount; (2-4) three latent smoke-media bugs newly exposed by the 08-08 smoke-hardening: genai import deferred to the torch-venv gate, LiteRT symbol check pointed at libtensorflowlite_c.so AND un-broken from the nm | grep -q SIGPIPE-under-pipefail false-stub, gst-libav gated on ffmpeg-executability; (5) libvulkan-dev Multi-Arch: same mirror skew on arm64 — cross-scoped dpkg --force-overwrite drop-in. Plus: parallel-arch OOM root-caused to INTRA-arch BuildKit DAG overcommit.

• Refactoring-backlog program: 6 agent sweep rounds + toolchain deep-dive + currency audit (36 legacy items proven already-fixed, 1 REGRESSION exposed: live buildkitd.toml lost gckeepstorage) -> lean batch-ordered backlog (docs/refactoring-backlog.md, journal archived), ~50 items executed same day. Tree now gated by 22 unit suites / 311 assertions (was 13/150), fix10 (PR100017 -nostdinc++ static guard), and two NEW gates: python-lint (ruff; gate=real-error classes) and secret-scan (gitleaks, enforcing, 2 public-trust-anchor FPs triaged into .gitleaksignore).

• Ops/infra: linux/host-config/ — canonical buildkitd.toml (gc 500GB + max-parallelism 4) + apply/verify scripts (the anti-regression answer); logs/ excluded from build contexts (2.38 GB/stage transfer, measured); services lane hardened (loopback bindings + lan override, compose pins + healthchecks, nginx security-header include validated via containerized nginx -t, benchmark runner atomic writes + JSONL persistence + the _manifest.json KeyError that killed every suite run’s final comparison); llm-stack contract tests wired into CI (ollama service + micro model).

• Docs: AGENTS.md Windows tables deduped into windows-builds.md (46 rows, 0 droppable, 1 real drift fixed), gate-list enumerations replaced by KNOWN_SLUGS pointers everywhere, GCC 16.1.0-era stragglers + TensorRT/cuDNN examples corrected, IREE + versions.env-toggle + smoke-deferral sections written, doc-literal scan added to sync_versions.py --check.

2026-08-10 (late) - sccache forensics round: poisoning excluded on BOTH cache levels, minimal repros clean, issue draft rewritten to the honest evidence state

• Owner challenge (“ist es wirklich upstream?”) audit: binary confirmed VANILLA mozilla/sccache main @ the #2722 merge (cargo install –git); SCCACHE_MULTILEVEL_CHAIN is an upstream feature (src/cache/multilevel.rs). Upstream prior art found: #1098/#1186 (our 10054 Windows crash family, open, no root cause), #2299 (arch-guard x host-preprocessor correctness bug, NVIDIA-reported, fixed pre-our-pin - proves the bug class), #2726.

• The audit EXPOSED an overclaim in our own draft: run 11 only had a fresh L0 - the WebDAV L2 was never reset. Follow-up killed that worry from the other side: the L2 store contains just 9 entries (all from micro-probes) - the chain’s multilevel write-through silently never fed WebDAV (separate puzzle, error-log dig pending post-run-13; until then cross-run CUDA caching rests on the L0 mount alone). Consequence: runs 10/11 compiled for REAL through the decomposition -> the miscompile happened at compile time, not via cache hits.

• Minimal wrapped-vs-bare discriminators (fresh disk-only cache, llvm-nm symbol diff, toolchain-image solves): define-guard and CUDA_ARCH-guard instantiation TUs x {plain, ORT-ish (4 gencode, -t2, -Xcompiler, extended-lambda), –options-file rsp} - ALL CLEAN. The instantiation loss requires real-ORT invocation complexity; next candidates -MD/-MF depgen, -forward-unknown-to-host-compiler, quoted rsp defines, concurrency. Next step: replay one affected TU’s exact generated command line (post-run-13).

• Issue draft + AGENTS failure row rewritten to exactly this evidence state (controls a-d, prior-art links, no overclaims). Post still awaiting owner go-ahead. Learned-the-hard-way note: nvcc response files are --options-file, not cl-style @file.

2026-08-10 (night) - run 12: ONNX vertex GREEN (bare-CUDA verdict confirmed); OpenCV died one layer deeper in MLAS -> patch 003; run 13 launched

• Run 12’s ONNX vertex went green at ~76 min INCLUDING the lld-link that killed runs 10/11 - the sccache-nvcc miscompile verdict is now triple-confirmed (run 5 bare green, 10/11 wrapped red, 12 bare green) and patches 004/005/006 are exonerated as link suspects. The vertex is committed as bk-windows-media-core-onnx; later solves FROM it.

• OpenCV vertex failed at 520 s in vendored MLAS - one layer past run 5’s failure (002’s force-include fix held): the bundled kernels are GAS/ELF .S only (.type sym,@function), and clang-cl IS a working GAS assembler, so the check_language(ASM) guard that protects MSVC does not fire - the integrated assembler then rejects ELF directives for a COFF target (“expected absolute expression”, SgemmKernelSse2.S). No MASM port exists upstream. 003-mlas-windows-skip.patch skips MLAS on Windows (dnn falls back to its built-in SGEMM, the same fallback upstream uses on Android; inference runs on ORT/DirectML anyway). Patch regenerated from a real git diff after a hand-written hunk mis-counted its header; Test-PatchesApplyClean green across all 11 mapped patches.

• The elevated buildkitd-env restore was DECLINED at the UAC prompt, so the new 0a gate got a targeted -SkipStepLogGate override (same philosophy as -SkipRdna4Gate: disk/shim gates stay armed) and run 13 launched with it

  • the 2 MiB step-log clip remains active until the restore happens between runs. Console logs now one timestamped file per launch (#30).

2026-08-10 (night, backlog execution part 2) - W1/W2 rest + 0b CI half: the Windows backlog is now closed except 27/28-analysis/31/0b-human

• #21: ArgQuoting AST detectors moved to modules/WindowsLint.Common.psm1; Invoke-Lint now runs them inside its single parse pass (violations FAIL the gate) over a fully recursive windows\ walk — which also brought windows\upstream into lint scope (141 files). The test suite keeps the positive controls only.

• #9/#29: probe-build-copy’s three lanes collapsed into one Invoke-ProbeLane runner; diagnostic tags adopted the diag- prefix (cleanup one-liner documented). Not yet live-smoked — one Test-BuildCopy.ps1 -Heavy after run 12’s media-core is the last check.

• #8/#30: Get-DiagnosticLogPath + Limit-DiagnosticLogs own the diagnostics’ log path/retention (driver keeps newest 80 stage logs, diagnostics newest 60).

• #14: verify-cuda-cache’s 21-statement RUN line became a COPY’d, lintable Test-Cache.ps1. #15: renamed to Sync-DefenderExclusions.ps1.

• #2 finish: reset-container-stores + verify-cuda-cache resolve tools via Get-PreferredToolPath; fixed reset-container-stores’ hardcoded D:\GitHub repo path and a latent Test-Path $null throw on missing buildctl.

• 0b (CI half): new patch-drift job in windows-scripts.yml runs Test-PatchesApplyClean on every windows/versions.env trigger; also repaired a pre-existing #requires line lodged INSIDE that script’s comment help. #32: answered — CI is hosted windows-latest, no containers, no AMD dGPU → RDNA4-immune; gate/toggle stay local-only.

• #28 groundwork: 15 s fleet sampler captured per-process WorkingSet/CPU through run 12’s ONNX compile (out\build-logs\onnx-tu-memory-samples-*.csv).

• Gates: lint 141/0/0 + AST-traps clean; tests 457/457.

2026-08-10 (night, backlog execution) - W3 media-closure batch + W2 preflight architecture + W1 quick wins, 457 tests green

• W3 (landed inside run 12’s already-busted closure window — the morning’s Dockerfile edits had already invalidated every downstream vertex, so these were cache-free): stall-guard v2 (#16: timed sccache --show-stats client probe is the verdict, CPU delta only a pre-filter; #20: no guard without a remote; #11: marker is the single print channel), attempt-scoped retry ladder (#17, +regression test; the run-10 “poisoned L0” paragraph RETRACTED — run 11’s fresh-mount control proved miscompile, not poisoning), $invokeNinja append flag removed (#10), Invoke-SourcePatchWithFallback with -Fatal 006 rung (#7+#19, all six build-onnx stanzas collapsed, 4 tests), Write-SccacheStatsToStderr called by all six source builds (#3), onnx ninja log persisted on the cache mount with one .prev generation (#4), SCCACHE_ERROR_LOG into C:\sccache\logs\ with early dir creation (#23).

• W2/0a: Assert-BuildkitdStepLogEnv — non-admin registry read that REFUSES to launch while the buildkitd service env lacks BUILDKIT_STEP_LOG_MAX_SIZE=-1 (the drift that hid verdicts all day); wired into build-buildkit.ps1. 0c: Driver.PreflightParity.Tests.ps1 owns lane parity (shared contract + reasoned allowlists + catch-all for new gates; caught Assert-ImageExists classification on first run).

• W1/#1+#6+#18+#5: RDNA4 hazard set single-sourced (Get-Rdna4HazardDevice) and the toggle primitive lifted into the module (Set-Rdna4DeviceState, post-state-verified); toggle script resolves all hazard SKUs by default, gained -NoPrompt/exit-1-on-failure; the A/B delegates its lanes to Test-BuildCopy.ps1 -Heavy; both drivers offer gate-specific -SkipRdna4Gate. #24: #requires -Version 7.0 across all 40 remaining test files, per-file EOL preserved.

• Gates after every sub-batch: Invoke-Lint 139 files 0/0, test suite grew 438 → 457 passed / 0 failed.

2026-08-10 (late night) - run 11 falsified L0 poisoning: the sccache nvcc path SILENTLY MISCOMPILES → CUDA launcher off, final for this pin

• Run 11 (fresh sccache-winamd64-2 mount) failed the providers_cuda link byte-identically to run 10 (QkvToContext<*, __nv_fp8_e4m3>, BiasSoftmaxImpl<double> undefined) — cache poisoning is falsified. The consistent explanation across all eleven runs: run 5 (bare nvcc) is the only run that ever linked green; runs 6/7 (wrapped) died before the link; 10/11 (wrapped) reached it first and lack the same arch-guarded instantiations. The pinned sccache’s nvcc decomposition drops device-conditional code — silent wrong code, disqualifying regardless of hit rate. (It also explains the Severity::k0 phantom: run 5 compiled triton_kernel.cu green with the WRONG 004 variant, bare — the macro collision only ever manifested through sccache preprocessing.)

• SCCACHE_NO_CUDA_LAUNCHER=1 is back in build-onnx — now as the FINAL state for this pin, with the three-canary re-enable bar (verify probe + fused_moe compile + full LINK canary) encoded at the call site, in the AGENTS failure row (rewritten from the falsified poisoning attribution), and in the upstream issue draft (now carrying both failure classes). Patch 006 stays: inert while unwrapped, load-bearing on any future retry.

• C/CXX caching remains on and proven; run 12 relaunched in the run-5 configuration plus all of today’s fixes (mlas patch, GenAI 0.15.2 ahead).

2026-08-10 (night) - backlog batch W1 round 2: #2 + #26 closed

• #2 closed with a scope correction: the “10+ single-candidate files” claim was an over-count (the grep matched files that already carry candidate lists). The truly single-candidate set was six sites — all now resolve the Program Files + D:\Stevedore candidates (reset-container-stores ×2, collect-host-docker-state, plus the three same-day diagnostics). deploy-shim-patch/verify-host-setup keep their overridable Program-Files param defaults deliberately: a shim DEPLOY target must not silently follow a stale D:\ layout (that path vanished on this host once before).

• #26 closed: both probe Dockerfiles take ARG BASE; Test-BuildCopy.ps1 pins it to versions.env’s WINDOWS_BASE_DIGEST (dependency-free parse — no module import in the first script a new host runs) on all three lanes. Smoke-verified live: probe green with probe base pinned: ...@sha256:d5bbb830…. The RDNA4 A/B keeps the tag default until #5 (probe reuse) lands.

• Gates green after each step (lint 137/0/0, tests 430/430).

2026-08-10 (night) - backlog batch W1 (host-only, zero cache impact, worked while run 11 compiled)

• Backlog items closed/advanced with gates green (lint 137/0/0, tests 430/430): #12 ArgQuoting manual case-loop → -notcontains; #13 fake-ninja double-guard folded; #22 verify-cuda-cache no longer mints a throwaway image per run (solve-only); #25 A/B verdict variables StrictMode-safe; #2 partial — the three 2026-08-10 diagnostics (probe-build-copy, verify-cuda-cache, test-rdna4-layer-lock) resolve buildctl/docker from the supported candidate list (Program Files + D:\Stevedore) instead of a single hardcoded path (7 legacy files remain); #6 partial — Set-Rdna4Gpu.ps1 -GpuName closes the wrong-SKU dead end (gate names a remedy that can now act on RX 9060/ R9700 hosts). Media-closure items (#10/#11/#16/#17) deliberately deferred until run 11 finishes — editing bind-mounted modules mid-chain would redefine later vertices in flight.

2026-08-10 (night) - run 10: link failure from POISONED sccache L0 → cache-mount id bumped

• Run 10 compiled all ~1891 ONNX objects but died at the DLL link with undefined template instantiations (QkvToContext<*, __nv_fp8_e4m3>, BiasSoftmaxImpl<double>). Root-cause hypothesis with the best fit: the persistent sccache L0 cache mount (id=sccache-winamd64) holds objects from runs 6/7, whose sccache server was KILLED mid-write (guard kills / manual unwedge) — truncated objects have been served as L0 hits ever since, and the mount survives every retry, so the failure could never self-heal. Fix: cache-mount id bumped to sccache-winamd64-2 in BOTH media Dockerfiles (fresh empty L0; the old one ages out via the 20GB/168h exec.cachemount GC tier). Non-admin, deterministic.

• Follow-up encoded in the backlog (#16/#17 guard redesign): a guard kill must be assumed to poison in-flight L0 writes — either verify the multilevel fork’s atomic-write behavior or purge/quarantine on kill.

• Note: the next run also picks up the owner’s GenAI v0.15.2 + LiteRT-LM 0.15.0 bumps already present in the Dockerfiles — backlog item 0b (bumps must ride the Windows lane) is live, not theoretical.

2026-08-10 (review) - 8-angle code review of windows/: 10 same-day fixes + refactor backlog

• Owner asked “is my chain clean code, what should I refactor” → systematic 8-finder review over windows/. Verified correctness findings were FIXED immediately (each with the failure it prevents): probe zero-lane false-green; Repair-WindowsComponentstore.ps1 still on the retired type=local probe shape (false-red on healthy hosts); classic lane build.ps1 missing the RDNA4 gate (isolation-probe verdict does not key on dGPU state); Get-SccacheStatsText not re-exported from WindowsSourceBuild.Common — the stderr stats line would have thrown CommandNotFound AFTER the multi-hour ONNX build (run 9 was stopped mid-flight over this); Dockerfile.heavy’s trailing-backslash COPY dest (Dockerfile escape char eats the newline); RDNA4 A/B now Tee’s full lane logs and VERIFIES the dGPU re-enable; hazard regex survives Radeon(TM)-style renames (test added, 430 total); opencv/001 patch EOL flip reverted (byte content is a layer-cache key); SCCACHE_ERROR_LOG parity for Dockerfile.media-merge-builder; stale launcher-opt-out comment in Invoke-CmakeConfigure rewritten to the shipped design.

• Everything not correctness-critical went to the new “Refactor Backlog (Windows container chain)” section at the end of this file’s sibling docs/windows-builds.md — 25 prioritized items (P1 drift/correctness-adjacent, P2 reuse, P3 hygiene), each with its failure scenario.

2026-08-10 (late) - sccache CUDA launcher: deterministic server crash scoped to the cuda_llm target → bare nvcc THERE, launcher (and CUDA caching) stays ON everywhere else

• Runs 6 and 7 died at ~4910 s (±2 s) on the same TU family (fused_moe_gemm_sm80_{bf16,f16}.generated.cu, os error 10054 on every client): the pinned source-built sccache’s nvcc decomposition crashes the server deterministically there — this is NOT the earlier 0%-CPU stall shape, and no retry ladder can pass a deterministic crash point. The identical run-over-run timeline (despite a hot L0) additionally shows the CUDA hit-rate at this commit is ~zero — the launcher cached nothing.

• NEW patches/onnxruntime/006-cuda-llm-bare-nvcc.patch (owner requirement: CUDA must cache — a blanket opt-out was rejected): the crash source is exactly ONE OBJECT library (onnxruntime_providers_cuda_llm, where all fused_moe_gemm_*.generated.cu live), so the patch clears that target’s CUDA_COMPILER_LAUNCHER property; every other CUDA target stays sccache-wrapped and cacheable. Applied with an inline-regex fallback. SCCACHE_NO_CUDA_LAUNCHER=1 remains only as the emergency full opt-out.

• sccache stats now go to STDERR after the ONNX build (sccache-stats| prefix, -Advanced) — stderr survives the 2 MiB step-log clip, so per-run CUDA hit/miss/error rates are finally visible (they were never observable before; the ~zero-hit-rate claim can now be confirmed or retired with data). Candidate-sccache re-enable criteria for the cuda_llm target stay encoded at the versions.env pin: Test-CudaCache.ps1 + an ONNX canary through the fused_moe launchers.

• Stall guard + full-speed retry ladder stay armed for the whole wrapped set.

• Never-swallow-logs sweep (owner directive): the buildkitd service env was found EMPTY — the repo-required BUILDKIT_STEP_LOG_MAX_SIZE=-1 (unlimited step logs, Install-NewHost.ps1 applies it) had been wiped by the 2026-08-09 Stevedore/repair work, and the default 2 MiB clip hid the stall-guard verdicts and sccache stats for three runs (re-apply + buildkitd restart pending the next between-runs window). SCCACHE_ERROR_LOG now persists inside the sccache cache mount (server postmortems survive the container). Test-BuildCopy.ps1 and Test-CudaCache.ps1 Tee their FULL lane output to out\build-logs\ and print the path (display keeps showing tails only). New AGENTS.md invariant bullet pins the principle.

2026-08-10 (resolution) - RUN-layer finalize 0x20: root cause is the ENABLED RDNA4 dGPU; build with it disabled

• Same-boot A/B on the RX 9070 XT host: disable the dGPU → tiny AND heavy RUN-layer finalize green on the first try; enable → red (upstream docker/for-win#14977 — AMD RDNA3.5/4 + Adrenalin lock freshly-written container layer files on process-isolated layer ops; issue open). Severity tracks the Windows patch level: pre-KB5101684 only heavyweight RUN layers tripped (light probes green — the host looked healthy while the chain died); post-KB5101684 (installed 2026-08-10) even 10-byte RUN layers fail. COPY-only layers are unaffected (no container involved). The full falsified list from the hunt: Defender (exclusions verified live; realtime-off blocked by tamper protection), WSearch/SysMain, daemon bounces, vmcompute restart, minifilter detaches (zero third-party filters on C:), --no-cache fresh IDs, settle delays, reboots, nanoserver base, split solves.

• Failed finalizes WEDGE hcs state: after one, even tiny RUN layers fail until a reboot (survives service restarts and vmcompute) — this cascade is why every earlier debugging round produced contradictory verdicts, and why the old probe (which crashed its own export on every run) kept hosts looking broken. The 2026-08-09 “Adrenaline reinstall fixed it” and “in-place repair fixed it” root-cause claims are SUPERSEDED: they coincided with patch-level/reboot changes that moved the trigger threshold.

• Lint-scope gap closed: Invoke-Lint.ps1 never covered windows\scripts\diagnostics\ — its scripts (GPU/isolation probes, now the RDNA4 A/B) were silently unlinted. Added recursively; 136 files, all clean.

• NEW diagnostic windows/scripts/diagnostics/Test-Rdna4LayerLock.ps1 (elevated): the RDNA4 A/B as a durable ~2-min tool — probes RUN-layer finalize with the dGPU enabled then disabled (finally-guarded re-enable), verdicts GONE / PRESENT / INCONCLUSIVE. Re-run after every Adrenalin/Windows update; GONE = retire the toggle workflow + gate.

• NEW preflight gate Assert-NoActiveRdna4Gpu (WindowsBuildDriver.Common, wired into build-buildkit.ps1 after Assert-ShimPatch; -SkipHostChecks overrides; 7 tests in BuildDriver.HostGates): refuses to start a chain that would die on its first RUN commit while an RDNA4 dGPU is enabled, and names the workflow — elevated Set-Rdna4Gpu.ps1 -Disable → build (display falls back to the iGPU; DirectML-on-host unavailable during the window) → re-enable. Set-Rdna4Gpu.ps1’s “obsolete” header is retracted.

• Chain relaunched with the dGPU disabled: base stage committed the previously-failing COPY and the VS Build Tools layer on the first attempt.

• Run 5 milestone: the ONNX vertex went GREEN end-to-end (compile, link, export, commit — first of the day; patches 001-005 all production-proven) and died one partition later in OpenCV 5.0.0’s newly bundled MLAS: NEW patches/opencv/002-mlas-clangcl-force-include.patch — mlas cmake treats clang-cl as GNU-Clang and passes -include + cstring, which the CL dialect parses as an input file; the patch adds an MSVC-frontend branch (/FIcstring + /w). 10/10 patches validate.

• NEW diagnostic windows/scripts/diagnostics/Test-CudaCache.ps1 — proves the sccache→nvcc(decomposed)→WebDAV path end to end: a tiny buildctl solve FROM the local toolchain image compiles one .cu twice; asserts the recompile HIT and that objects reached the store. First run verified 2026-08-10: 4/4 component hits (CUDA/Device/PTX/CUBIN), 4 objects + check-marker written to the dufs root (which was EMPTY before — no prior run had ever populated the CUDA cache). Traps encoded: sccache’s nvcc parser needs the -ccbin= (equals) form, and bare in-container compiles need INCLUDE assembled from the MSVC root + Windows SDK (the images bake no VS env).

• sccache’s nvcc launcher DEADLOCKED mid-ONNX (run 4): NEW Start-SccacheStallGuard watchdog in Invoke-NinjaBuildWithRetry (WindowsSourceBuild.Common) — CUDA caching STAYS ON (owner requirement). Symptom: sccache server + 9 clients all at 0 % CPU, zero backend connections (dufs healthy), compile crawl/stall for 40 min; killing the server alone is NOT enough (clients block forever on the dead pipe). The guard samples the compiler fleet every 60 s from a background job; three consecutive ~zero-CPU samples with sccache alive → kill ALL sccache processes → in-flight compiles fail fast → the retry ladder resumes with cache hits for everything already compiled. Guard is inert when sccache is not on PATH (unit tests unaffected). Emergency escape: SCCACHE_NO_CUDA_LAUNCHER=1 (honored by Invoke-CMakeConfigure) drops only the CUDA launcher, C/CXX caching stays. Run-6 hardening: the deadlock recurs (~40-80 min apart) and a guard kill is NOT OOM-shaped, so the single -j<RetryJobs> retry lost the run (in-flight compiles die with os error 10054, and the guard’s console verdict sat beyond the 2 MiB step-log clip). The guard now appends every kill to a marker file IMMEDIATELY, and Invoke-NinjaBuildWithRetry gives marker-confirmed guard-kill failures up to 3 FULL--j retries (compiled objects are L0 hits) before the incremental attempt. 2 new tests pin the ladder (429 total). Also learned: C:\sccache (the L0 tier) is a persistent BuildKit cache mount (id=sccache-winamd64) — L0 hits make no WebDAV traffic, so a quiet L2 store during a mostly-cached vertex is normal, not a defect (the multilevel disk,webdav write-through itself was probe-verified against the live endpoint).

• NEW patches/onnxruntime/005-xqa-host-stub-sccache.patch — ORT 1.28.0’s XQA (paged-attention) kernels are the only ORT sources with host/device- divergent include guards (host pass keys on the cmake define HAS_SM80_OR_LATER); sccache’s nvcc decomposition can drop target -D defines in the host sub-step, so the host namespace lacks smemSize/kernelType/cacheVTileSeqLen and the generated stub dies with C2039/C2065 in x_?.cudafe1.stub.c (the synthetic x_?.cu name is the sccache fingerprint). Since this build pins CUDA_ARCHITECTURES=80;86;89;90, the patch emits the host stub unconditionally (documented as not-upstreamable-as-is). Found at ~1295 s of run 3 — one whack-a-mole past the tunable.h fix, which run 3 proved good (triton_kernel.cu compiled).

• Two media-lane source patches for the 2026-08-03 version bumps (first Windows build attempt since): NEW patches/onnxruntime/004-tunable-severity-macro-collision.patch — ORT v1.28.0 + CUDA 13.3: wingdi.h’s classic #define ERROR 0 (reached despite -DNOGDI when a header in triton_kernel.h’s chain includes wingdi directly) pre-expands through the LOGS_DEFAULT forwarding macro into the nonexistent Severity::k0 in tunable.h (nvcc: enum ... has no member "k0" at the LOGS_DEFAULT(ERROR) line, first TU triton_kernel.cu); fixed with guarded #undef ERROR/#undef VERBOSE after the header’s includes, applied in Build-OnnxFromSource.ps1’s CUDA branch with an inline-regex fallback. (First fix attempt undef’d only VERBOSE — a summarizer misreported which LOGS_DEFAULT(...) sat on the error line; the moved-but-identical error exposed it. Read the line, not the expectation.) REGENERATED patches/opencv/001-cmake-clang-cl-compat.patch against OpenCV 5.0.0 — the old hunk context still named the CMP0218 policy block that 5.0.0 removed, so the patch (applied with NO fallback) would have thrown an hour into media-core. Test-PatchesApplyClean.ps1 (which reads pins from versions.env): all 8 patches OK.

2026-08-10 - build probe was lying: two pwsh bugs fixed (probe verdicts trustworthy again; the same day’s RESOLUTION entry above tells where “green” actually ended)

• Test-BuildCopy.ps1 carried two bugs since its introduction that made a healthy BuildKit lane read as broken. (1) The unquoted --output type=local,dest=$outDir reached buildctl as VERBATIM SOURCE TEXT — pwsh passes a bareword comma-attribute argument to native commands with no variable expansion (measured on pwsh 7.6.4) — so the export wrote into a directory literally named $outDir and then died in the receiver. (2) The docker-classic lane assigned $docker = "...\docker.exe", which IS the [switch]$Docker parameter variable (names are case-insensitive), throwing Cannot convert ... String to ... SwitchParameter — that lane had never executed at all.

• The probe’s BK lane now exports type=image,name=docker.io/local/ kataglyphis:probe-build-copy,unpack=true — the same output path build-buildkit.ps1 uses — so a green probe covers layer commit AND the finalize/export/unpack reimport. It also exits non-zero naming each failing lane (previously it could end green-looking with a failed lane). Never use a type=local export of a Windows image as a health signal: the local exporter dies client-side (error from receiver: write ...\Boot\Fonts\ <font>.ttf: file already closed) even on a healthy host — new AGENTS.md Common Failure Modes row.

• Verdict with the fixed probe on the discovered host (RX 9070 XT/26200): light lanes green on buildkit (commit + export + unpack) — but NOT chain-green. The real chain’s first COPY after the heavy pwsh-install RUN dies deterministically (ActivateLayer 0x20 at child finalize/reimport, FRESH snapshot IDs under -NoCache — not poisoned cache). Minimal repro isolated and committed as the probe’s new -Heavy lane (Dockerfile.heavy: RUN writing 2×100 MB, then a one-file COPY — fails in ~1 min while a plain nested multi-file COPY on servercore commits fine): the child’s finalize fails while the fresh heavy parent layer stays held. Ruled out live: Defender (full exclusion set verified, incl. MsMpEng), daemon state (elevated containerd+buildkitd bounce), poisoned cache (fresh IDs under --no-cache), and time (a 90 s settle layer’s own commit fails identically) — host-level hcs/filter hold. Next levers: reboot → -Heavy re-probe → clean Adrenaline reinstall → shim A/B → healthy host (Dockerfile.heavy doubles as a 60 s upstream repro). docker-classic legacy COPY still fails there (mkdir \\?\Volume{...}\C:.). Chain launches also fail-fast at preflight when the dufs sccache endpoint is down — its ONLOGON task does not re-fire after repair reboots; Start-ScheduledTask dufs-sccache cures it (done today). README / host-setup / AGENTS.md updated.

• NEW windows/scripts/tests/Native.ArgQuoting.Tests.ps1 (8 tests, in the pre-build gate): AST sweep of every windows/ script for both trap classes — bareword comma-attribute args referencing variables, and non-boolean assignments to [switch] parameter variables in their own scope — plus positive controls pinning the detectors themselves. Both sweeps are clean repo-wide after the probe fix. New AGENTS.md § Windows Build Invariants bullet documents the traps.

2026-08-09 (resolution) - Windows COPY-commit failure: root cause was a faulty AMD Adrenaline install, NOT the RDNA GPU — SUPERSEDED by the 2026-08-10 resolution above (the enabled RDNA4 dGPU IS the holder; this entry is kept as history)

• The ‘AMD RDNA3.5/RDNA4 GPU breaks Windows-container layer commits’ hypothesis (upstream microsoft/Windows-Containers#623) is RETRACTED as a root cause (corrected 2026-08-09). On the discovered host (Ryzen 9 9950X

  • RX 9070 XT, 25H2/26200) the build-COPY failure - hcsshim::ActivateLayer 0x20 on buildkit and mkdir \\?\Volume{<GUID>}\C:. on docker legacy, both lanes, every COPY layer, surviving -NoCache, restarts, Defender exclusions, a full store reset and a reboot - was caused by a FAULTY AMD ADRENALINE installation. A clean reinstall fixed it (probe: windows/scripts/diagnostics/Test-BuildCopy.ps1). GPU-disable never cured it; Set-Rdna4Gpu.ps1 is now obsolete as a fix. The Linux cross lane and all repo gates were never affected. Docs updated (superseding the night entry below): AGENTS.md Common Failure Modes + script table, docs/windows-host-setup.md gate, docs/windows-builds.md diagnostic section, README.md.

2026-08-09 - LLM stack: GPU mode override + Qwen3-Coder deploy

• NEW linux/llm-stack/docker-compose.gpu.yml — a compose overlay that gives the Ollama service all NVIDIA GPUs (deploy.resources.reservations.devices, driver nvidia) and raises the default context via OLLAMA_CONTEXT_LENGTH. The base docker-compose.yml stays CPU-only; GPU hosts run docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d after installing the nvidia-container-toolkit (install commands in the llm-stack README). OLLAMA_KV_CACHE_TYPE=q8_0 and flash attention remain the defaults.

• Deployed qwen3-coder:30b (30.5B A3B MoE, Q4_K_M, ~18 GB) on a 2× NVIDIA host (12 GB + 16 GB): 100 % GPU placement (9.9 GB / 12.4 GB per card), warm decode ~137 tok/s, ~47 s cold load, context tuned to 64K. Perf measured via the /api/generate metrics.

• Docs: linux/llm-stack/README.md gained a GPU-mode section (toolkit install + override usage + ollama ps GPU check) and a VRAM/context sizing table (~104 KB/token q8_0 KV ⇒ 28 GB total = ~64K cap; 256K needs >45 GB). The “CPU-only inference” architecture note is now “CPU-only by default, GPU optional”. Repo README.md and AGENTS.md Quick Reference now point to the LLM stack (previously undiscoverable from either).

2026-08-09 (night) - Windows lane on a 25H2 host: platform COPY regression found; sccache source build proven host-side

The base->final GPU verification run that motivated the sccache source build hit a wall that turned out to be the HOST OS, not the repo.

• SUPERSEDED 2026-08-09 — root cause corrected above: FAULTY AMD ADRENALINE install (reinstall fixed it), NOT the RDNA GPU. Historical entry kept verbatim below. Originally claimed: Windows 11 build 26200 (25H2 line) COPY-into-layer failure — isolated to a AMD RDNA3.5/RDNA4 GPU defect (upstream microsoft/Windows-Containers#623), NOT a blanket build break and NOT the ISO (corrected 2026-08-09).** buildkitd: failed to reimport snapshot: hcsshim::ActivateLayer 0x20 (file used by another process), deterministic across fresh chain-IDs, survives -NoCache, service restarts, vmwp kills, Defender exclusions, a full store reset AND a reboot. docker-classic fallback: mkdir \?\Volume{}\C:. invalid directory name, under process AND Hyper-V isolation. A minimal 3-layer probe isolates it: FROM servercore + RUN commits fine, the first COPY layer never does. Root cause is the AMD GPU: on this box (Ryzen 9 9950X + Radeon RX 9070 XT, RDNA4) hcsshim layer VHD mounting breaks at COPY commit; the upstream issue’s own matrix proves RDNA3.5/RDNA4 = broken while RDNA1/RDNA2 iGPU/Intel = fine. NOT the OS: a same-build 26200 machine builds fine, sfc/DISM report 0 components corrupt, and even projfs.sys is absent on the working machine too. Fix (reproducer-verified): disable the AMD GPU in Device Manager (RDP survives via the Remote Display Adapter). On the discovered host (2026-08-09) GPU-disable did NOT cure it - real upstream, persisted here; root cause on that box remains undetermined FINAL-verdict-update: FINAL 2026-08-09: BK build lane unusable on that host UPDATED after the in-place Windows repair (same build 26200): the layer-COMMIT wall is FIXED (all layers incl. existing-dir writes commit on buildkit); only the final EXPORT reimport 0x20 persists, and the Defender engine is unkillable by design so that lever stays untestable - residual host hcs export behavior. The full base->final chain is one export step from green on this host. - the reimport 0x20 persists on buildkit 0.32.0 AND a throwaway v0.32.2 (instrumented A/B), i.e. host-level, not engine-version; classic lane (dockerd) builds the same shapes fine there; heavy/GPU chain goes on the working host.. Docs updated: docs/windows-host-setup.md OS gate + AGENTS.md Common Failure Modes carry the corrected row. The Linux cross lane and all repo gates are unaffected.

• sccache source build verified HOST-SIDE (the part that needs no image): the pinned SCCACHE_GIT_REV = e9b15a3 is confirmed from upstream to BE the mozilla/sccache#2722 merge (“Fix nvcc dryrun parsing for CUDA 13.3”, carries test_group_nvcc_subcommands_with_simt_only_cicc_input); the EXACT command Install-RustToolchain.ps1 runs (cargo install sccache --locked --git https://github.com/mozilla/sccache --rev <rev>) compiles, links and installs cleanly (exit 0, 3m25s, sccache.exe in CARGO_BIN, --version = 0.17.0 exactly as the commit documented). The wiring itself remains covered by the 412-test / 0-lint gates (verify-toolchain CARGO_BIN assert, CMAKE_CUDA_COMPILER_LAUNCHER). The one thing still pending is a real ONNX CUDA kernel cache-hit in an image build - which needs a supporting (non-25H2) host.

• New-host bring-up (Install-NewHost.ps1 + verify-host-setup fix + magic-constant purge, this morning’s entry) proved out on this fresh host: verify-host-setup all-green, patched shim deployed and hash-recorded, CNI confs on the live subnet, dufs L2 up with logon task. Host-side probe toolchain (rustup gnu via Strawberry’s bundled mingw linker) installed for the verification above - throwaway, not part of any image.

2026-08-09 (late) - Windows lane: one-script new-host bring-up + magic-constant purge + verify crash fix

Verified live while bringing up a brand-new host (this one) for the sccache source-build verification run; every fix below is what a fresh Stevedore box actually trips over.

• NEW windows/scripts/host/Install-NewHost.ps1 - the scriptable half of docs/windows-host-setup.md Phases A5+C as ONE elevated, idempotent run (-ReportOnly safe, refuses while a build is live): authors the CNI .conflist from the LIVE vEthernet (nat) subnet (derived network/prefix+GW at runtime - the magic subnet literals are gone from the docs), then orchestrates the canonical per-concern scripts (apply-containerd-config -> .conf derive + debug flags + teardown env + Defender; apply-buildkitd-gcpolicy

  • the BUILDKIT_STEP_LOG_* env; deploy-shim-patch - BUILDING the 45min/100min fixed-constant shim from hcsshim source when no -ShimPath is given, Go via scoop; dufs install/start/ONLOGON-task/machine endpoint env). Every sub-script is called with a HASHTABLE splat - the first draft used array splats and hit the documented position-binding trap in -ReportOnly (-ReportOnly arriving as $ServiceName), exactly the AGENTS array-splat rule.

• Test-HostSetup.ps1 line-212 crash fixed: (Get-ItemProperty ...).Environment on a service whose Environment value does not exist threw PropertyNotFound, which under StrictMode surfaced as an unset-variable error and CRASHED the script mid-run on the common drifted host - silently skipping the teardown-env and debug-flag checks and under-counting the verdict (reported 2 warnings instead of 3). Registry values that do not exist now degrade to honest WARNs (.PSObject.Properties.Name -contains reader), matching the Defender check’s degrade-to-UNKNOWN contract.

• Docs purge of stale example values: docs/windows-host-setup.md A5 and docs/windows-builds.md section Getting it going no longer hand out the reference host’s 172.31.32.0/20 subnet as copy-paste gospel - both now say “derive” and point at Install-NewHost.ps1; README’s fresh-machine pointer gains the one-run path.

2026-08-09 (early) — Linux lane: validator split + a locale bug the split’s own probe caught

• validate_compiler_for_target decomposed (complexity item 9): the 108-line monolith is now five independently-callable checks (cc_check{dumpmachine,binary_elf,object,loader,qemu_exec}) with explicit args — deliberately NOT the VCS* implicit-global convention.

• Locale bug found by the split’s verification probe: every readelf parser in the tree matched the ENGLISH “Machine:”/”interpreter:” strings — on a non-C-locale host (this one is de_DE) readelf localizes (“Maschine:”) and the ELF-machine checks failed spuriously; inside containers (LC=C) the bug was invisible but latent. All parsed readelf -h/-l/-d calls across 7 files now run under LC_ALL=C. Host probe: 6/6 PASS.

2026-08-08 (night) — Linux lane: complexity F-G + TVM shallow/commit-pin clone

• setup-package-image’s 102-line dual-purpose function split (audit F-G): select_dev_packages / install_dev_packages / clang_embedded_deb_version (the previously function-nested, globally-leaking _clang_ver) / pin_clang_alternatives — the old name remains as its caller’s 4-line sequence. All load-bearing comments preserved.

• TVM clone hardened: shallow --depth 1 --branch <tag> instead of the old bare recursive clone (which pulled the whole default branch + all submodules unpinned before ever checking out), plus a TVM_COMMIT opt-in pin (commit beats tag; v0.25.0’s commit recorded in versions.env) — completing the *_COMMIT convention for the second compiler-class clone.

• Repo hygiene sweep came back clean: no build artifacts tracked, .gitignore already covers resource-monitor outputs and out/.

2026-08-08 (night) — Linux lane: supply-chain round 3 — build executors frozen

• Round-3 completion: Vulkan SDK and GStreamer-Android-universal tarball sha pins landed once their streamed hashes finished — the audit’s entire class-(d) (completely unverified fetch) list is now EMPTY.

• The ~20-site unpinned uv pip install surface is closed for every binary-shaping package: meson, ninja, cython, pybind11, setuptools, wheel, scikit-build-core, setuptools-scm — and auditwheel/patchelf, which REWRITE the shipped wheels’ ELF headers — are pinned via a new PY_*_VERSION family in versions.env (frozen to what builds were already resolving; the torch wheelhouse keeps its documented setuptools<82 ceiling via a dedicated pin). Pure-python data deps stay floating by design. Wired at 8 install sites with the inline-default safety-net convention.

• abseil headers now fetch the immutable /archive/<commit>.tar.gz form, sha-verified (ABSEIL_COMMIT + ABSEIL_TARBALL_SHA256) — the tag-tarball form is movable and not byte-stable.

• verify-parity.sh main() decomposed (complexity F-F, zero blast radius): five responsibilities → five functions matching the file’s own check_* shape; behavior verified (–help, missing-args, exit codes).

• AGENTS.md: supply-chain discipline codified in Linux Build Rules; a 16.1.0/16.2.0 GCC version drift in the rules text fixed.

2026-08-08 (night) — Linux lane: supply-chain hardening round 2 — pins for every trust anchor

All computed from upstream (streamed sha256, gzip/manifest-verified) and wired with the warn-if-unset pattern:

• CUDA apt keyring deb pinned (x86_64 + sbsa) — the apt trust anchor for the whole NVIDIA lane was fetched with no hash; ROCm apt signing key pinned (was TOFU via wget|gpg); Android commandlinetools zip pinned (sdkmanager bootstraps the NDK cross compiler from it); Flutter SDK pinned with Google’s official sha256 from releases_linux.json; rustup/uv installer pins populated (the fail-closed mechanism existed with EMPTY keys — both effectively ran unverified remote scripts as root).

• TensorFlow C SDK: the pin was fiction. Upstream stopped publishing libtensorflow C builds after 2.18.0 (x86_64 only) — the pinned 2.21.0 never existed as an artifact, the GitHub URL has 404’d since 2.19, and a 2>/dev/null swallowed it: ffmpeg silently shipped WITHOUT its TF DNN backend while versions.env claimed otherwise. Now: 2.18.0 from the GCS mirror, sha-pinned, aarch64 says loudly that no upstream build exists, and the do-not-bump constraint is documented in versions.env.

2026-08-08 (night) — Linux lane: supply-chain round 1 + safe complexity refactors

(See commit 68bc11e: TLS redirect/wget hardening, CPython fail-closed, NodeSource/npmjs pipe-to-shell removal, npm ci lockfile-exact, verified Kitware/LLVM host-repo helpers, cerbero tag pinning, meson wrap-update removal; lib/ prelude drift + clang-extractor + cross-fallback parity with two new test suites.)

2026-08-08 (night) — Linux lane: smoke-depth round (audit round 3, lens 2)

A capability×depth audit of every smoke layer found the deepest ML coverage living in an EXTERNAL repo’s app smoke — torch, torchvision, onnxruntime inference and OpenCV imencode had zero in-repo functional coverage, and several capabilities had none anywhere. Presence checks upgraded to real execution (all device-less, no network, seconds each):

• GStreamer mandatory plugins (libav/opencv/onnx/tflite) now GATE in the runtime smoke on the real target arch under qemu, and in smoke-media’s native branch — a present-but-unloadable plugin (the observed webrtcbin2/gtk4 class) was previously only a WARN-count. Plus a data roundtrip (videoconvert!jpegenc → 4 real JPEG frames) beyond the registry-only fakesink pipeline.

• Python stdlib battery (ssl/sqlite3/lzma/bz2/zlib/hashlib/ctypes, exercised not just imported) in smoke-toolchain — the textbook from-source CPython failure; _sqlite3 was checked NOWHERE in linux/ and joins build_python.sh’s staging warn-list too.

• torch forward+backward, torchvision nms/._C + v2.Resize, and a real onnxruntime InferenceSession (model generated in-process via torch.onnx.export — no fabricated bytes) in smoke-torch-venv, gated STV_COMPUTE=1 (default on).

• ffmpeg codec depth: buildconf-vs-registration consistency for x265/dav1d/svtav1/vpx/opus (build-ffmpeg probe-gates –enable-*, so a silently-missed probe DROPS a codec while the build stays green) + real encode/decode roundtrips for libx265/libvpx-vp9 — all inside the binary-executes guard.

• Cross-compiler loader assertion: the emitted ELF’s PT_INTERP must request the TARGET’s dynamic loader (wrong-sysroot links succeed and only die on target); opportunistic static-binary qemu-run (exit-42 proof) when qemu-user is present.

• Rust: version pinned against RUST_VERSION (the old check asserted “rustc” appears in rustc --version — could never fail), host compile+RUN, and per-target emit-obj (rustup lists targets whose std rlibs are missing). node/npm: first coverage at all (version pin + JS execution) — the LiteRT-web WASM gate silently self-disables without node.

• LiteRT: nm -D symbol check (TfLiteInterpreterCreate/TfLiteModelCreate) — works on foreign-arch ELF, so the cross branch gets it too; a 12-byte stub used to pass [ -f ]. nvcc: device-less global kernel compile + the fail-open hole closed (ENABLE_NVIDIA=true with no nvcc now fails). GenAI: shipped-but-unimportable is now FAIL, absent stays INFO. Vulkan runtime: real vkEnumerateInstanceVersion call (works with zero ICDs; a healthy loader cannot fail it). Android NDK: the compile smoke its header promised since creation (per-target object + ELF-machine assertion; the NDK clang is a host binary, runs on every branch).

• Fail-open holes closed: absent cross-Python staging dir now FAILS for requested foreign arches (ran zero checks before); smoke-media’s hardcoded “torch not installed” INFO replaced with a real venv probe (it was false in the package image); smoke-vulkan’s vkvia | head rc swallow fixed.

2026-08-08 (night) — Linux lane: orphan sweep (audit round 3, lens 1)

A dedicated dead-weight audit (things wired to nothing), with every dynamic script-dispatch mechanism enumerated first so glob/variable-built paths could not produce false orphans:

• Real arm64 bug: wasm-opt.sh built the aarch64 binaryen asset name but versions.env pinned no BINARYEN_LINUX_AARCH64_SHA256 — the download died “No pinned SHA256” on any arm64 host. Pinned (sha computed from the upstream version_131 tarball, gzip verified).

• Dead legacy alias TVM_VULKAN_KEEP_SDK_LIBS removed from vulkan.sh (sole occurrence repo-wide; the canonical knob is VULKAN_KEEP_SDK_LIBS).

• Legacy flutter shims (setup-flutter-{arm64,x86-64}.sh, kept for external ExternalLib consumers) no longer carry a hardcoded 3.44.8 fallback that had already drifted from the 3.44.9 pin — they now require the versions.env value they load anyway.

• Deleted: 02-toolchain/rust/Build-Linux.sh (zero references AND a duplicate re-implementation of its five cargo_* siblings) and 06-packaging/package_archive.sh (zero references, zero docs; its Windows “twin” is equally unreferenced, so no parity obligation).

• Documented as consumer surface (shipped into images, invoked by external repos, previously invisible): lib/{ctest-run,docs-build, rust-toolchain}.sh (now in the AGENTS.md + linux-build-basics.md inventories alongside their already-documented siblings), 02-toolchain/rust/cargo_*, 02-toolchain/python/ci_*.sh, 01-core/setup-host-deps.sh.

• make lint-workflows target added (preflight ran it; the Makefile only exposed lint-dockerfiles).

• Clean bill elsewhere: zero dead Dockerfile ARG/ENVs, zero unreferenced patches/data files (verify-patch-integrity already gates patches), zero dead Makefile targets/workflow steps.

2026-08-08 (night) — Linux lane: audit round 2 leftovers cleared

The verified-but-queued remainder of the four audit reports:

• Orphan verify branches wired (Dockerfile.media): onnxruntime-gpu runs after the GPU EP build (GPU-enabled images had NO gate on those artifacts) and armnn runs after the arm64 ACL/ArmNN build (a failed build left empty /opt/armnn + /opt/acl shipping unexamined). Both branches had existed caller-less since creation.

• make verify-chain can now fail: STALE links are counted and the explicit --verify-chain path exits 2 when any exist — it used to warn and exit 0, an explicit verification that could not fail. The automatic partial-run protection (_chain_assert_ancestry, hard-fail) is unchanged. Makefile help text updated.

• verify-parity.sh judges by rc, not sentinel: || echo "FAILED" appended AFTER the captured traceback, so the first-word parse read “Traceback” and an ImportError could never increment failures.

• Runtime Vulkan check three-way: ctypes.CDLL of the loader does NOT need an ICD/GPU — a load failure means the library is missing/broken, and the runtime image always ships the Vulkan runtime files. Missing/unloadable is now a FAIL; only container-infra errors stay WARN.

2026-08-08 (night) — Linux lane: audit round 2, Klasse C — test gaps closed

• Zero-assertion suites now FAIL: t_summary treats _T_RUN=0 as a failure (a gutted suite used to print “0 assertion(s) passed” and stay green), and run-tests.sh aggregates the per-suite counts — the final line now reads “N suites, M assertions”, so a coverage collapse is visible in every log.

• Four new/extended suites (7→11 suites, 68→120 assertions):

  • test-parallelism.sh — pins mem_capped_jobs’ RAM/cores formula, the ≥1 floor, and the new PARALLEL_JOBS validation (fix included: the override was emitted unvalidated — PARALLEL_JOBS=0 went straight into ninja -j0; now rejected with a warning).

  • test-cli-parsers.sh — pins the new central two-arg value guard (fix included in dispatch_parsed_args: a trailing --target-arches or one that swallowed the next flag used to assign “”/”–push” silently and fall through to CROSS_DEFAULT_ARCHES — building all three arches).

  • test-stage-defs.sh — the REAL cross_stage_tag/cross_build_mem_divisor/ graph-validation (test-disk-guard/test-ancestry stub these; until now no test executed the real ones). Also asserts the Klasse-B ENABLE_NVIDIA forwarding.

  • test-smoke-arch-parity.sh — asserts smoke-common.sh’s inline fallback maps agree with the canonical platform.sh/arch-mapping.sh for all three arches (this file had already caused two silent-skip bugs).

  • Extended: riscv64→RISCV LLVM backend assert (a tempting “consistency rename” would break the compiler stage), runtime_artifact_platform/ _image_ref (wrong-arch artifact COPY class), version-forwarding negative asserts (an unset var must NOT become --build-arg FOO= overriding the Dockerfile default with empty; # noforward must hold).

• IFS bug class killed by construction: arch_list_to_words and smoke_arch_words now emit NEWLINE-separated words — for x in $(...) splits under both the default and the strict IFS=$'\n\t', retiring the 16 latent for-loop sites the audit found (all consumers verified compatible: for-loops, wc -w, unquoted argv). Parity suite pins the property.

• Gates that could not fail, now real: smoke-torch-venv fails when STV_REQUIRE_VENV=1 and /opt/venv is absent (the package wrapper-smoke sets it — the venv gate used to SKIP+pass exactly when setup-torch-venv failed hardest); the SDK image asserts a non-empty /opt/vulkan and (except riscv64) an executable /opt/flutter/bin/flutter (both shipped with ZERO verification); smoke-torch-venv reports TVM presence/version per-arch with EXP_TVM as opt-in hard pin (Dockerfile.media’s comment claimed an import tvm runtime gate that never existed — comment corrected); cross_stage_validate_graph (pure, sub-second) now runs in preflight (slug stage-graph) instead of only at build kickoff.

2026-08-08 (night) — Linux lane: audit round 2, Klasse D — convention bugs

• USE_CCACHE/USE_SCCACHE/USE_LLD now accept both truthiness spellings (0/false/no/off, any case). Previously only the literal string "false" disabled them — USE_CCACHE=0, the convention half the fleet uses (and what ENABLE_SCCACHE_* expects), was silently ignored. Fixed in compiler-cache.sh (+ shared _flag_disabled helper), cmake-cache-linker.sh, build-ffmpeg.sh, onnxruntime lib/common.sh, ffmpeg-probe-framework.sh (inline case in the standalone-bundled files, per bundling policy). Verified live: USE_CCACHE=0 now prints “ccache disabled”.

• Bare sudo in the ONNX AMD/NVIDIA steps (30-build-native-amd.sh, 30-build-native-nvidia.sh) replaced with the CPU sibling’s guarded pattern (command -v probe + SKIP_DEP_INSTALL honor) — they died rc 127 on images without sudo while the CPU step degraded gracefully.

• common.sh’s sudo fallback now sets SUDO_WRAP too (it set only SUDO; a ${SUDO_WRAP} consumer reaching that path aborted under set -u).

• verify-arg-consistency.sh no longer mixes WARN: and WARNING: prefixes.

2026-08-08 (night) — Linux lane: audit round 2, Klasse B — contract drift

• CUDA_ARCHITECTURES carried literal quote characters into CMake: the only quoted value in versions.env, and load_versions_env exports values verbatim — CMake received "80 and 90", and the documented Hopper 90→90a suffix transform was a silent no-op (the string ends in "). Value dequoted in versions.env (the file format is unquoted inert data, as the loader header documents) AND the loader now strips one pair of surrounding quotes — defense in depth for the class. Verified live: the transform now yields …;90a.

• Android OpenCV built a different OpenCV than the chain: OPENCV_VERSION="${1:-5.x}" — the dispatcher passes no arguments and the env was ignored, so every android image cloned the MOVING 5.x branch while the chain ships tag 5.0.0. Now env-first with the 5.0.0 inline default (sibling pattern), and Dockerfile.media’s final stage exports ENV OPENCV_VERSION so the android stages inherit the pin the same way they already inherit GSTREAMER_VERSION.

• Dockerfile.media still fell back to /opt/gcc-16.1.0 in three places — a path that no longer exists since the 16.2.0 bump (every script-side fallback had been bumped; the Dockerfile inlines were invisible to verify-arg-consistency, which only parses ARG NAME= lines).

• Stale nested fallbacks (all invisible to the checker’s $-containing literal guard): GenAI v0.15.0→v0.15.2, VVdec v3.1.0→v3.2.0, Python 3.14.6→3.14.7 ×2 (build_python.sh would have died on the 3.14.7 checksum with a misleading error) + the same stale 3.14.6 on the Windows side (Build-OpencvFromSource.ps1).

• ENABLE_NVIDIA/ENABLE_AMD now reach the cross lane: the runtime lane honored them, cross_stage_build_args dropped them — a GPU-configured runtime could sit on CPU-only media artifacts with no warning. Forwarded (only when set) for the media stage.

• install_vulkan_sdk zero-arg call aborted “unbound variable”: the fallback referenced setup-dependencies.sh’s flag-local VULKAN_VERSION_DEFAULT; the chain now ends in the canonical VULKAN_VERSION pin.

• Four comments whose stated contract had drifted from the code they sit on (abseil default, three GCC-16.1.0 claims) corrected.

2026-08-08 (night) — Linux lane: audit round 2, Klasse A — error-path masking

Four-perspective audit (error paths, contracts, tests, conventions); this entry is the error-path class. Every fix below closes a path where a real failure was reported as success:

• ELF wrong-arch gate was dead code (validate-media-runtime.sh): the clean-scan exit 0 sat BEFORE the ELF architecture validation; NEEDED sonames resolve by name, so a wrong-arch binary scanned clean and the gate never ran. Now an else branch — ELF validation runs on every path.

• Stale-rootfs export on failed builds: _build_one_artifact (build-runtime-artifacts.sh) and _sdk_arch_build (build-sdk-artifacts.sh) ran under if !-suppressed errexit with no || return 1 — a failed runtime_build_chain/cross_stage_run fell through to export_rootfs_from_image, which exported the previous run’s tag and reported green. Guards added.

• parallel-loop lost workers that die via exit: err() terminates the background subshell before || touch failed-flag runs, and the join discarded wait’s rc — a dead lane read as green under PARALLEL_ARCHS=1. A nested ( ) layer now absorbs the exit into a return code.

• app-wheels gate was vacuous: the Dockerfile’s .placeholder (written exactly when the wheelhouse build failed) satisfied “dir not empty”. The riscv64 verify now requires a real *.whl; ALLOW_EMPTY_APP_WHEELS=1 is the explicit escape hatch.

• verify-media-artifacts could not fail for litert / genai / opencv-core: litert checked /usr/local/{include,lib} (already filled by base’s CPython); genai checked dirs its own producer mkdir -p’s on every skip path; opencv-core used INFO-only optional checks. All three now require stage-specific artifacts (new side-effect-free probe_lib/verify_any_lib helpers also fix the verify_A || verify_B idiom that counted A’s failure even when B passed). genai verify mirrors its producer’s legitimate cross-build skip instead of “verifying” pre-created empty dirs.

• smoke-media masking: a non-executable ffmpeg/gst-inspect was a PASS (“will work at runtime”) — now INFO + ELF-magic assertion, with the functional gate named; the OpenCV cvtColor roundtrip and GStreamer pipeline checks had no else-branch (could not fail) — now fail when execution demonstrably works; ffmpeg encode failure fails when the binary executes and advertises libx264; onnxruntime import-failure now proves the library exists instead of claiming presence unchecked.

• verify-cuda-stack.sh rewritten: || true) pasted inside three command substitutions made the “not found” branches unreachable and hard-failed healthy images under set -e. Now honest warn-only (stderr, no -e) with CUDA_STACK_STRICT=1 as the real gate for complete-stack images.

• base-image bootstrap fails fast: the apt-update retry loop breaked away its terminal failure and continued; ca-certificates install failure was a log line. A broken mirror/CA store now aborts the bootstrap with the culprit named instead of surfacing hours later as an opaque TLS error.

2026-08-08 (evening) — Linux lane: IREE tblgen Exec-format failure, and the binfmt registration that silently died

media-arm64 failed: IREE’s NATIVE tblgen was cross-compiled

The app-wheelhouse IREE cross build died with Exec format error on llvm-project/NATIVE/bin/llvm-min-tblgen: LLVM’s CrossCompile.cmake defaults the NATIVE sub-build’s compilers to the outer cross compilers, so the tblgen that must run on the amd64 build host was built for arm64. Fix: build-app-wheelhouse.sh now passes -DCROSS_TOOLCHAIN_FLAGS_NATIVE pinning the true host compilers (+ ccache launchers — closing the nested-sub-build caching item from the backlog in the same stroke). riscv64 never hit this only because qemu binfmt silently emulated the wrong-arch tblgen — slowly.

Which exposed: binfmt registrations die on containerd restart

The morning’s shim-failure systemctl --user restart containerd silently wiped the rootlesskit-namespace qemu registrations (they are namespace-lifetime, not host-lifetime). Re-registered for arm64+riscv64 and installed the rootless-binfmt.service –user unit so login/boot re-registers automatically. Two bugs fixed in setup-rootless-binfmt.sh itself along the way: it claimed “pulling” but never pulled (image save fails “not found” on a fresh host), and its blob-detection pipeline tar -tf | grep -q self-destructed under pipefail (SIGPIPE — shell bug class 2) so extraction skipped every blob. AGENTS.md’s two stale recommendations of the non-working rootless tonistiigi/binfmt --install container corrected to the helper script.

The foreign chain marked arm64 failed and moved on to media-riscv64 (by design); media-arm64 + android-arm64 re-run after the chain with the fix.

2026-08-08 (evening) — Linux lane: structural round 1 (cache-key closure + drift bugs + dead code)

Dockerfile.base no longer cache-keys on ~120 files (A1)

base’s six RUNs bind-mounted all of 01-core + 02-toolchain, so editing any of ~120 files — including host-only orchestrator modules the image never executes — busted base and cascaded a rebuild through the entire chain, undoing the toolchain stage’s careful per-file mount lists one tier up. The mounts are now the traced 15-file transitive closure of base-image.sh (mirror RUN: 2 files). Validated with a real from-scratch base build; the first attempt failed exit 127 because the static trace missed use-fast-ubuntu-mirror.sh, which bootstrap-ca execs rather than sources — closure tracing must follow exec/bash edges, not just source lines. Marginal cost of a future 01-core edit drops from “full chain rebuild” to near zero.

Three drift bugs between intentional clones (A2, A4, A5)

• cross_build_is_active existed 5× with 3 semantics; the documented arch-normalization fix had reached 1 of 5 copies. The raw copies compared OCI names against uname -m machine names and reported “cross active” on native arm64 hosts. All fallbacks now normalize via arch_normalize.

• compiler-resolution.sh never shipped to the android stages, so IREE-android’s fallback command -v gcc resolved the custom cross GCC from the inherited toolchain PATH as its host compiler (live bug, masked by best-effort gating). Dockerfile.android now COPYs the canonical script; the fallback prefers explicit /usr/bin compilers like the litert copy.

• Dockerfile.package hand-rolled the ports-mirror sed: it ignored the USE_FAST_UBUNTU_MIRROR gate and never derived ports-from-archive. It now runs the canonical use-fast-ubuntu-mirror.sh.

Dead code out (B1–B4, B6)

smoke-wrapper.sh (orphaned; two docs falsely claimed wrapper-smoke runs it — both corrected), create_deb() (104 lines, zero callers, positioned after the script’s outputs were written), run_quiet() (zero callers), the vestigial smoke-runtime-image.sh COPY in Dockerfile.package, and AGENTS.md’s claim of a media_build_preamble_init alias that does not exist.

Deferred to the post-chain batch (they touch files the running foreign chain bind-mounts): A3 parse-table, A6 dir-walker unification, B5 smoke-runtime decomposition. Full status in docs/refactoring-backlog.md.

2026-08-08 (afternoon) — Linux lane: the –no-push hole, and a forensic audit of “green”

--no-push full-chain handoffs never worked on this host

The BuildKit OCI worker keeps its own image store: nerdctl build -t loads results into containerd, which the next build’s FROM never consults — the mutable parent tag resolves against the registry. Every downstream stage of a --no-push chain silently built on the last PUSHED parent. Proof: a freshly built compiler shipped /opt/gcc-16.2.0 while the sdk built “from” it contained /opt/gcc-16.1.0 (a months-old ghcr image); FROM repo@<containerd-digest> errors “not found”. Two full validation runs were lost to this before the digest trail exposed it. Interim: push mode is the only correct full-chain flow (docs updated, a warning fires at --no-push parse time); real fix (OCI-layout build-context handoff, the runtime lane’s existing mechanism) is specced in the backlog. The manifest is protected by running --to-stage android + the runtime lane with --skip-manifest until all arches exist.

A fifth shell bug class, found by auditing the helper scripts

Functions whose last statement is [ cond ] && action return 1 in the healthy case; under set -e the guard tool kills itself. The flagship victim: verify-critical-fixes.sh — the script guarding the Five Critical Fixes — aborted after fix 1 whenever the fixes were actually present, silently skipping fixes 2–9 and the summary. It only ever looked green because hosts don’t carry the staged payloads. Fixed there, in smoke-common.sh (the “Unknown arch” guard was unreachable), and lint/preflight/NVIDIA-lane guards in the same sweep.

Forensic audit of the build logs: what “21/21 PASS” was hiding

Two-agent sweep over ~12 MB of logs, each claim re-verified:

• Built 0.15.2, shipped 0.14.0: the app’s uv.lock outvoted the chain’s freshly built onnxruntime-genai wheel (--find-links only OFFERS). Fixed (pre-install + --no-install-package), and genai added to the version-pin smoke so this class cannot recur unnoticed.

• The libcamera pin was undercut at the finish line: the media validator did not know meson’s lib/<multiarch> install dirs, declared the build’s own libs missing, and apt-installed Ubuntu’s older libcamera as a shadow copy — a false-positive “repair”. Fixed (multiarch dirs in the scan path).

• ccache delivers zero in LLVM’s nested sub-builds (189 identical objects compiled twice, second pass 1.9% slower): CROSS_TOOLCHAIN_FLAGS_NATIVE forwards no compiler launcher. Plus: no ccache statistics are emitted anywhere — and the 2MiB step-log clip truncates stdout only, so stats (and anything that must survive) belong on stderr. Both queued.

• Still open, prioritized in the backlog: TVM ships import-broken (built against distro LLVM 22.1.2, not the pinned 22.1.8 — the telltale line prints at INFO), pyav is pinned but never installed, FFmpeg’s TF/OpenVINO DNN backends died silently, OpenCV links distro GStreamer/FFmpeg due to stage ordering, three assertion-free fallback-PASSes, inner smoke warnings invisible to the outer verdict.

• Verified NON-issues (do not chase): “missing NVENC” is ENABLE_NVIDIA-gated by design; runtime Python 3.14.4 is Ubuntu’s distro CPython as venv base (a decision, not a stale layer); “is not a commit!” clone warnings are annotated-tag peeling.

Also — periphery audit (workflows, hooks, tooling)

A dedicated sweep over the never-audited edges. The flagship: the pre-commit hook’s sphinx gate pointed at a nonexistent docs/source/ with the real error swallowed — every commit on a hook-enabled clone failed; fixing the path then exposed 12 real docs warnings under -W (10 orphaned pages now in an “Operations & Reference” toctree, 2 unknown-lexer blocks) — the strict gate is green end-to-end for the first time. Also fixed: deps.json’s libcamera entry now bound to LIBCAMERA_VERSION (the public license pages published “git master” past yesterday’s pin), the install-deps action’s dirname-of-empty-string putting . on GITHUB_PATH, a benchmark-viewer build that reported success over a failed npm build, three unreachable FAIL branches in the webserver flutter smoke, least-privilege permissions: on the two unpinned workflows, SHA-pins for the last two mutable action refs, rename coverage (--diff-filter=ACMR) and a loud git-grep failure mode in the hook, and the license generator now fails on unknown versions.env vars and defaults to check mode. Remaining periphery items are in the backlog.

Also

• Owner priorities codified: AGENTS.md § Project priorities (speed AND stability AND tests, docs always in the same work unit), README § Engineering principles.

• buildkitd GC budget pinned (~/.config/buildkit/buildkitd.toml, gckeepstorage 500GB) + step-log-size drop-in (both effective at the next between-runs restart). Unexplained: base cache-missed after the first restart despite unchanged mounts and a surviving store — under investigation before cross-restart layer reuse is trusted.

• versions.env: 11 bumps (checksums from official sources), libcamera pinned for the first time (v0.7.2 — the only media library that tracked master), OpenCV moved to the immutable 5.0.0 tag; ROCm deliberately HELD (new “TheRock” releases 404 on the old apt path); LiteRT-LM 0.15.0 verified to keep the protobuf 6.31.1 coupling.

• latest-cross-amd64 shipped: full from-base chain in push mode with digest-pinned handoffs + live ancestry annotations; runtime smoke 21/21 after a host containerd-shim failure was diagnosed (every nerdctl run died; builds unaffected) and the services restarted. The ancestry guard and the annotation-based --verify-chain verdicts had their first real-world successes the same day.

2026-08-08 — Windows lane: backlog cleared before the from-toolchain rebuild

The remaining four items, closed so the chain restarts against a tree with no known open work. Two of them turned out to be blocked only by a third.

• Warning floods cut at the source. 16 % of a chain log (72 864 of 459 061 lines) was four upstream constructs repeated thousands of times, which matters because buildkitd clips a RUN step’s log at 2 MiB and then deadlocks it. Targeted suppressions, never a blanket -w: -Wno-deprecated-copy (OpenCV matx.hpp), /clang:-Wno-unused-value (ONNX), -Wno-documentation-unknown-command (TVM), and — because STL4037 is emitted by the MSVC STL headers themselves and no clang group can switch it off — _SILENCE_NONFLOATING_COMPLEX_DEPRECATION_WARNING for IREE/MLIR, at directory scope where it survives LLVM’s HandleLLVMOptions stripping. OpenCV’s is safe for the CUDA path because the repo’s own patch strips -W* before nvcc’s cl.exe host compiler sees it (verified against the patch, not the comment above it). New Measure-BuildWarnings.ps1 reports each family against its pre-suppression baseline, so the next run PROVES each flag still earns its place rather than it becoming folklore.

• Smoke test split, 1 573 → 1 386 lines, harness into WindowsSmokeTest.Common.psm1 (no Dockerfile change — the final image already COPYs the whole modules dir). The move had one hazard and both halves of it fail silently: Assert-Test read $ExitOnFirstFailure out of the calling script’s scope, which a module cannot see (the switch would have quietly stopped working), and the summary read $script:passed, which across a module boundary would have reported 0 passed / 0 failed and exited 0 on any run. Both are explicit state now. An AST inventory of every assertion call site is 194 before and 194 after, identical as a set; 11 new tests, suite at 412.

• Pre-commit hooks are enabled, and the reason they were not is gone: preflight.sh used bare python3, which on this host is the Microsoft Store stub, so every commit would have failed for reasons unrelated to the commit. It now probes for an interpreter that can actually execute code.

• Submodule pin drift (Kataglyphis-DocumANTation, UV_VERSION 0.12.1 vs 0.12.3) kept the version-snapshot check red, which is what blocked the hooks. Fixed and committed in that submodule; it still needs a push there plus a pointer bump here, left explicit because it is a different repository.

2026-08-08 — Windows lane: three gaps that could not fail loudly

Landed in the window a concurrent versions.env pin bump opened: PYTHON_VERSION 3.14.7 invalidates Dockerfile.base from its COPY versions.env down and every media stage under the toolchain, so these changes cost no rebuild that was not already owed. Each one is the same shape as the rest of this week’s work — a check that was structurally incapable of reporting the failure it existed for.

• The FFmpeg .pc gate could not fail in its worst case. It sat inside if (Test-Path $ffPkgConfigDir), so a missing lib\pkgconfig — the most complete failure available — skipped every assertion without a word. Extracted to Assert-FfmpegPkgConfig, which treats an absent directory as fatal, and called outside that guard. Five unit tests, one per failure mode, reaching the function by AST extraction so it stays out of the three media branches’ compile closure (the reason Remove-MakefileShowIncludes moved out of the shared module on 2026-08-03).

• The base image’s PATH had two dead entries and was missing a live one. SCOOP_HOME/SCOOP_GLOBAL are scoop app roots and hold no executables. Meanwhile flutter is installed --global, so C:\ProgramData\scoop\shims exists and was on no PATH entry at all — a 2026-07-14 comment had removed it as a “never-created dir”, which stopped being true the moment anything was installed globally. Invisible only because FLUTTER_BIN is baked separately; any future global package would have been unresolvable by name. Restored (user shims keep priority) and asserted by the smoke test.

• An unresolved merge conflict had been committed into docs/refactoring-backlog.md and survived several commits: two sections were appended concurrently and never merged. Markdown and shell lint both pass a conflict marker, because it is valid text. Resolved (content verified identical modulo the markers), and .githooks/pre-commit now greps the staged content for <<<<<<< / >>>>>>> — verified to fire on the exact commit that carried the bug. A bare ======= is deliberately not matched: it is a legitimate Markdown setext underline, and real conflicts carry the others.

• Nothing was linting the git hooks. Writing that guard surfaced a live SC1072/SC1073 parse error already sitting in .githooks/pre-commit: a comment starting with the word “shellcheck” reads as a malformed directive and aborts ShellCheck’s parse of the whole file. It survived because lint-shell.sh filters to *.sh and a git hook cannot carry that suffix. Comment reworded, and lint-shell.sh now also accepts explicitly-passed extension-less files with a shell shebang — default sweep unchanged (223 files), staged-file coverage gained, and the hook now lints itself.

Note: core.hooksPath is unset on the primary dev clone, so none of these hooks have been running there — the documented one-time git config core.hooksPath .githooks (AGENTS.md) is still pending, and is left to the owner because it changes commit behaviour for every process writing to that tree.

2026-08-08 — Linux cross lane: four bug classes, machine-checked ancestry, live caching

Driven by a from-base amd64 rebuild of :latest-cross, fixing every failure as the chain hit it. The theme mirrors yesterday’s: silent failure made loud.

Four bash bug classes found live, fixed repo-wide, and lint-gated

1. trap … RETURN leaks to the caller — a RETURN trap set inside a function fires again when the CALLER returns, where the function’s locals are gone; under set -u this killed build-cross-chain.sh AFTER every stage had succeeded. Three instances (parallel-loop, context-management).

2. Unguarded pipelines under set -euo pipefail — du on a not-yet-created cache dir aborted the orchestrator on FIRST runs; readelf on linker scripts, dpkg -S on unowned files, find | head SIGPIPE and friends would have killed the media validators mid-stage. ~10 sites.

3. Comma-split loops break under IFS=$'\n\t' — for x in ${list//,/ } runs ONCE with the whole list as one bogus element in strict-IFS scripts. Broke the GCC GPG key import AND would have killed the compiler stage’s multi-target Python staging. New lint suite (test-ifs-safety.sh) bans the idiom; safe pattern is IFS=',' read -r -a.

4. Vendor scripts sourced under set -u — LunarG’s Vulkan setup-env.sh reads $1 unguarded and is sourced with explicitly cleared args; killed the sdk stage’s TVM step. Vendor sourcing now suspends nounset and restores it.

Cross-invocation ancestry is machine-checked now

Digest pinning only ever protected a SINGLE run. Every pushed cross stage now records its parent’s digest as an OCI manifest annotation (org.kataglyphis.parent-digest); partial runs (--from-stage after base) walk the recorded chain against the registry and hard-fail on a stale ancestor (linux/scripts/01-core/ancestry.sh). --verify-chain gives real FRESH/STALE verdicts from the same annotations. The old “after a compiler push start from sdk” rule is enforced by the machine, not the reader.

The GCC GPG failure was a key-rotation, not tampering

gcc-16.2.0 is signed by Richard Biener’s key; the script pinned only Jakub Jelinek’s and reported the missing public key as possible tampering (with SHA512 already OK). Now: accepted key SET, verdict via gpg --status-fd (NO_PUBKEY → warn/skip per policy; BADSIG → fatal), signer checked against the set so an arbitrary imported key cannot pass.

Toolchain caching went from decorative to real

The ccache wiring was inverted: the GCC RUN mounted the cache but never exec’d ccache; the LLVM RUN exec’d ccache but never mounted the cache. Fixed both, plus CCACHE_BASEDIR/SLOPPINESS (without which per-target build dirs made identical TUs never hit) and a multi-word-CC PATH fix for the Canadian path. --with-build-config=bootstrap-ccache was PROPOSED and REJECTED — GCC 16 ships no such config (verified against the tarball). Per-target GCC builds can now run concurrently behind GCC_PARALLEL_TARGETS=1 (serial apt pre-pass, divided JOBS, per-target logs; default off).

Version pins: complete and current

versions.env audited for completeness (libcamera was the ONLY unpinned media library — now LIBCAMERA_VERSION=v0.7.2, and the generated wheel stops lying about its version) and currency (11 bumps incl. Python 3.14.7, Node 26.7.0, OpenCV 5.x→5.0.0 tag = last non-reproducible media pin closed; ROCm deliberately HELD — the new upstream releases 404 on the old apt path). All checksums fetched from official sources; a new checker section catches the case-mapped version literals the ARG checks could not see.

Also

• --no-push validation runs no longer validate STALE registry images: the wrapper smoke pulled the published tag over the freshly built one, and the runtime handoff pulled cross-android-<arch> over the local build (BUILT_THIS_RUN now set on the local path too). Runtime chain failures propagate instead of reporting success.

• Disk preflight measures the cache dir’s own filesystem (not its parent’s) and survives first runs; --final-image is no longer silently overridden.

• apt.llvm.org 5xx no longer kills a multi-hour layer (falls back to source).

• Regression suites: test-ancestry.sh, test-parallel-loop.sh, test-ifs-safety.sh — auto-discovered by the pre-commit script-tests gate.

2026-08-07 — Windows lane: reproducibility, mandatory plugins, honest gates

The theme is less the repairs than what they have in common: several things had been failing silently for months, so most of this work is about making failure loud and early.

Mandatory GStreamer plugins are a contract now

libav, opencv, onnx and tflite were absent from the published winamd64 image and nothing was ever red — meson’s auto feature state means skip silently, and the healthcheck printed [PASS] for plugins that did not exist. Four unrelated root causes, diagnosed against gstreamer 1.29.2:

• opencv — OpenCV ships no .pc at all (confirmed: zero files in the built image). One is now authored, enumerating the import libraries from the real install (64 of them) instead of a hand-kept list that would rot.

• onnx — ONNX Runtime ships no .pc on any platform; one is emitted.

• libav — subprojects/FFmpeg.wrap provides the libav* modules pinned to FFmpeg 7.1.1, and -Dwrap_mode=forcefallback forced meson onto it, so pkg-config was never consulted: the build fetched a second, older FFmpeg instead of the n9.0 it had just built. The wrap is disabled before configure.

• tflite — consults no pkg-config at all. It probes the compiler for tensorflow/lite/c/c_api.h, the pre-rename path, while LiteRT ships the post-rename tflite/ layout; an alias tree is staged. Confirmed in the field that upstream’s first library name (tensorflowlite_c) does not exist here — only its fallback tensorflow-lite.

The set lives in Get-RequiredGstPlugin and is enforced at four points that previously disagreed: a pkg-config pre-flight (checking version floors, not just presence), meson features set to enabled, a post-install gst-inspect gate that throws, and smoke-test assertions. tensorfilter is deliberately excluded — it is an NNStreamer element this repo never builds.

FFmpeg’s .pc files were unusable

Found by probing the built image rather than waiting for the merge stage: Version: .. (configure found neither a VERSION file nor git tags, because the source is a GitHub auto-tarball) and MSYS-style prefix=/c/… paths that clang-cl cannot resolve. The empty version alone kept gst-libav out, independently of the wrap. Both fixed, and gated at the end of the FFmpeg stage.

Reproducibility

• LLVM, ninja and nasm pinned (LLVM_WINDOWS_VERSION, …). The OS base was digest-pinned while the very next layer installed whatever scoop served that day — and five patches in this tree are written against a specific clang-cl. Asserted at base-build time.

• C:\toolchain-manifest.json records every pinned input as a pin/resolved pair plus the floating ones, so which compiler built this image is answerable from the artifact. It captures the MSVC toolset (14.51.36231) that floats inside VS major 18 and was previously recorded nowhere.

• versions.env no longer invalidates the whole media chain. It was COPY’d into the stage all three branches descend from, so three Windows-only pins re-ran all six media compiles (~90 min of ONNX among them). Versions now travel as build-args; the file is demoted to a gap-filler by a precedence rule that distinguishes a real build-arg override from a value merely inherited from the base image’s machine environment.

Gates that stop lying

• Disk is checked per stage, with floors calibrated against measured consumption, on every drive the build uses (not just C:), in both lanes.

• The runhcs shim is identified by the SHA256 recorded at install time instead of by file size; Publish-ShimPatch.ps1 -RecordCurrent arms that without a redeploy.

• The CNI conf must exist as BOTH .conf and .conflist — buildkitd reads one, nerdctl the other, and “converting” between them cost a launched chain. The .conf is now derived from the .conflist.

• Retries stop immediately when a failure repeats byte-for-byte (a poisoned snapshot, whose remedy is -NoCache on that stage) but still retry snapshot-mount contention, which repeats verbatim and clears anyway. The merge stage’s -MaxAttempts 5 had been dead code, because its failure signature was never in the transient pattern.

2026-07-30 — Agentic loop: backlog-driven planner skip + completed-task pruning

• Skip planner when tasks are pending (backlog.skipPlannerWhenTasksPending, default true): while BACKLOG.md still has unchecked tasks, iterations go straight to the executor instead of paying for a planning pass.

• Completed tasks are deleted from the backlog (backlog.deleteCompletedTasks, default true): executor prompts now instruct deleting the finished entry (summary goes into the commit message), and a deterministic pruner (remove_checked_tasks / Remove-CheckedBacklogTasks) removes any lingering - [x]/- [X] blocks (title + indented body) before each auto-commit and at drain start. Completed work stays visible in git history instead of growing the file.

2026-07-30 — Agentic loop: live streaming output

• Claude engine streams by default: claude -p now runs with --output-format stream-json --verbose (config engines.claude.streamOutput, default true) and both libraries render the events live to console + log: session start, one line per tool call, assistant text per turn, tool errors, and a final turns / duration / cost summary. Set streamOutput: false to return to the silent text mode.

• Bash opencode invocation streams too: output is echoed line-by-line to console + log as it arrives instead of being buffered until exit (the PowerShell module already streamed).

2026-07-30 — Agentic loop: Claude Code engine + robustness

• Engine abstraction in both agentic-loop libraries (linux/scripts/lib/agentic-loop.sh, windows/scripts/modules/WindowsAgenticLoop.Common.psm1): opencode and claude (Claude Code CLI, headless claude -p) backends behind a single dispatcher (invoke_agent / Invoke-AgenticAgent). Selection via config engine, AGENTIC_ENGINE, or runner flag; model overrides via AGENTIC_PLANNER_MODEL / AGENTIC_EXECUTOR_MODEL.

• Claude engine: role system prompts via --append-system-prompt-file, planner sandboxed with --allowed-tools, executor permission mode configurable (default bypassPermissions), planner --fallback-model support.

• Robustness: retry with linear backoff per agent invocation, per-role timeouts (plannerTimeoutSeconds / executorTimeoutSeconds), build-failure fixer phase (executor-tier model gets the build log tail, then one rebuild), consecutive-build-failure cap that stops the loop, dry-run stall guard.

• Shared loop features moved into the libraries: planner-only / executor-only modes, max-iteration override, default phase prompts.

• PowerShell module: new exports Resolve-AgenticEngine, Invoke-ClaudeCode, Invoke-AgenticAgent, Invoke-AgentProcess, Invoke-BuildFixer, Get-AgenticConfigValue, Get-AgentTimeoutForRole; module version 1.1.0. Fixed the refactor planning cycle erroneously reusing the executor prompt.

• Pester suite extended to 38 tests (engine resolution, dispatcher, claude dry-run, engine-override loop smoke tests).